From 1fcb38143cd8b54a3d74965a6caaba98c5685d7d Mon Sep 17 00:00:00 2001 From: Glenn Jocher Date: Tue, 12 May 2026 01:26:31 +0800 Subject: [PATCH] Update 301 redirects --- .github/ISSUE_TEMPLATE/bug-report.yml | 2 +- .github/ISSUE_TEMPLATE/feature-request.yml | 2 +- README.md | 8 ++-- docs/en/compare/damo-yolo-vs-efficientdet.md | 14 +++---- docs/en/compare/damo-yolo-vs-pp-yoloe.md | 10 ++--- docs/en/compare/damo-yolo-vs-rtdetr.md | 14 +++---- docs/en/compare/damo-yolo-vs-yolo11.md | 24 +++++------ docs/en/compare/damo-yolo-vs-yolo26.md | 18 ++++----- docs/en/compare/damo-yolo-vs-yolov10.md | 24 +++++------ docs/en/compare/damo-yolo-vs-yolov5.md | 22 +++++----- docs/en/compare/damo-yolo-vs-yolov6.md | 26 ++++++------ docs/en/compare/damo-yolo-vs-yolov7.md | 16 ++++---- docs/en/compare/damo-yolo-vs-yolov8.md | 10 ++--- docs/en/compare/damo-yolo-vs-yolov9.md | 12 +++--- docs/en/compare/damo-yolo-vs-yolox.md | 6 +-- docs/en/compare/efficientdet-vs-damo-yolo.md | 22 +++++----- docs/en/compare/efficientdet-vs-pp-yoloe.md | 4 +- docs/en/compare/efficientdet-vs-rtdetr.md | 22 +++++----- docs/en/compare/efficientdet-vs-yolo11.md | 16 ++++---- docs/en/compare/efficientdet-vs-yolo26.md | 20 +++++----- docs/en/compare/efficientdet-vs-yolov10.md | 18 ++++----- docs/en/compare/efficientdet-vs-yolov5.md | 16 ++++---- docs/en/compare/efficientdet-vs-yolov6.md | 20 +++++----- docs/en/compare/efficientdet-vs-yolov7.md | 22 +++++----- docs/en/compare/efficientdet-vs-yolov8.md | 14 +++---- docs/en/compare/efficientdet-vs-yolov9.md | 20 +++++----- docs/en/compare/efficientdet-vs-yolox.md | 14 +++---- docs/en/compare/index.md | 6 +-- docs/en/compare/pp-yoloe-vs-damo-yolo.md | 4 +- docs/en/compare/pp-yoloe-vs-efficientdet.md | 20 +++++----- docs/en/compare/pp-yoloe-vs-rtdetr.md | 10 ++--- docs/en/compare/pp-yoloe-vs-yolo11.md | 42 ++++++++++---------- docs/en/compare/pp-yoloe-vs-yolo26.md | 20 +++++----- docs/en/compare/pp-yoloe-vs-yolov10.md | 22 +++++----- docs/en/compare/pp-yoloe-vs-yolov5.md | 14 +++---- docs/en/compare/pp-yoloe-vs-yolov6.md | 22 +++++----- docs/en/compare/pp-yoloe-vs-yolov7.md | 16 ++++---- docs/en/compare/pp-yoloe-vs-yolov8.md | 28 ++++++------- docs/en/compare/pp-yoloe-vs-yolov9.md | 24 +++++------ docs/en/compare/pp-yoloe-vs-yolox.md | 14 +++---- docs/en/compare/rtdetr-vs-damo-yolo.md | 14 +++---- docs/en/compare/rtdetr-vs-efficientdet.md | 24 +++++------ docs/en/compare/rtdetr-vs-pp-yoloe.md | 16 ++++---- docs/en/compare/rtdetr-vs-yolo11.md | 22 +++++----- docs/en/compare/rtdetr-vs-yolo26.md | 20 +++++----- docs/en/compare/rtdetr-vs-yolov10.md | 18 ++++----- docs/en/compare/rtdetr-vs-yolov5.md | 14 +++---- docs/en/compare/rtdetr-vs-yolov6.md | 20 +++++----- docs/en/compare/rtdetr-vs-yolov7.md | 22 +++++----- docs/en/compare/rtdetr-vs-yolov8.md | 22 +++++----- docs/en/compare/rtdetr-vs-yolov9.md | 12 +++--- docs/en/compare/rtdetr-vs-yolox.md | 6 +-- docs/en/compare/yolo11-vs-damo-yolo.md | 22 +++++----- docs/en/compare/yolo11-vs-efficientdet.md | 24 +++++------ docs/en/compare/yolo11-vs-pp-yoloe.md | 18 ++++----- docs/en/compare/yolo11-vs-rtdetr.md | 32 +++++++-------- docs/en/compare/yolo11-vs-yolo26.md | 18 ++++----- docs/en/compare/yolo11-vs-yolov10.md | 30 +++++++------- docs/en/compare/yolo11-vs-yolov5.md | 10 ++--- docs/en/compare/yolo11-vs-yolov6.md | 20 +++++----- docs/en/compare/yolo11-vs-yolov7.md | 12 +++--- docs/en/compare/yolo11-vs-yolov8.md | 24 +++++------ docs/en/compare/yolo11-vs-yolov9.md | 12 +++--- docs/en/compare/yolo11-vs-yolox.md | 30 +++++++------- docs/en/compare/yolo26-vs-damo-yolo.md | 18 ++++----- docs/en/compare/yolo26-vs-efficientdet.md | 14 +++---- docs/en/compare/yolo26-vs-pp-yoloe.md | 16 ++++---- docs/en/compare/yolo26-vs-rtdetr.md | 24 +++++------ docs/en/compare/yolo26-vs-yolo11.md | 22 +++++----- docs/en/compare/yolo26-vs-yolov10.md | 18 ++++----- docs/en/compare/yolo26-vs-yolov5.md | 10 ++--- docs/en/compare/yolo26-vs-yolov6.md | 20 +++++----- docs/en/compare/yolo26-vs-yolov7.md | 12 +++--- docs/en/compare/yolo26-vs-yolov8.md | 28 ++++++------- docs/en/compare/yolo26-vs-yolov9.md | 18 ++++----- docs/en/compare/yolo26-vs-yolox.md | 18 ++++----- docs/en/compare/yolov10-vs-damo-yolo.md | 22 +++++----- docs/en/compare/yolov10-vs-efficientdet.md | 22 +++++----- docs/en/compare/yolov10-vs-pp-yoloe.md | 22 +++++----- docs/en/compare/yolov10-vs-rtdetr.md | 22 +++++----- docs/en/compare/yolov10-vs-yolo11.md | 28 ++++++------- docs/en/compare/yolov10-vs-yolo26.md | 14 +++---- docs/en/compare/yolov10-vs-yolov5.md | 20 +++++----- docs/en/compare/yolov10-vs-yolov6.md | 26 ++++++------ docs/en/compare/yolov10-vs-yolov7.md | 30 +++++++------- docs/en/compare/yolov10-vs-yolov8.md | 12 +++--- docs/en/compare/yolov10-vs-yolov9.md | 22 +++++----- docs/en/compare/yolov10-vs-yolox.md | 22 +++++----- docs/en/compare/yolov5-vs-damo-yolo.md | 24 +++++------ docs/en/compare/yolov5-vs-efficientdet.md | 12 +++--- docs/en/compare/yolov5-vs-pp-yoloe.md | 22 +++++----- docs/en/compare/yolov5-vs-rtdetr.md | 24 +++++------ docs/en/compare/yolov5-vs-yolo11.md | 18 ++++----- docs/en/compare/yolov5-vs-yolo26.md | 16 ++++---- docs/en/compare/yolov5-vs-yolov10.md | 22 +++++----- docs/en/compare/yolov5-vs-yolov6.md | 18 ++++----- docs/en/compare/yolov5-vs-yolov7.md | 14 +++---- docs/en/compare/yolov5-vs-yolov8.md | 20 +++++----- docs/en/compare/yolov5-vs-yolov9.md | 22 +++++----- docs/en/compare/yolov5-vs-yolox.md | 12 +++--- docs/en/compare/yolov6-vs-damo-yolo.md | 16 ++++---- docs/en/compare/yolov6-vs-efficientdet.md | 18 ++++----- docs/en/compare/yolov6-vs-pp-yoloe.md | 26 ++++++------ docs/en/compare/yolov6-vs-rtdetr.md | 18 ++++----- docs/en/compare/yolov6-vs-yolo11.md | 32 +++++++-------- docs/en/compare/yolov6-vs-yolo26.md | 18 ++++----- docs/en/compare/yolov6-vs-yolov10.md | 18 ++++----- docs/en/compare/yolov6-vs-yolov5.md | 8 ++-- docs/en/compare/yolov6-vs-yolov7.md | 18 ++++----- docs/en/compare/yolov6-vs-yolov8.md | 26 ++++++------ docs/en/compare/yolov6-vs-yolov9.md | 14 +++---- docs/en/compare/yolov6-vs-yolox.md | 16 ++++---- docs/en/compare/yolov7-vs-damo-yolo.md | 16 ++++---- docs/en/compare/yolov7-vs-efficientdet.md | 24 +++++------ docs/en/compare/yolov7-vs-pp-yoloe.md | 12 +++--- docs/en/compare/yolov7-vs-rtdetr.md | 16 ++++---- docs/en/compare/yolov7-vs-yolo11.md | 16 ++++---- docs/en/compare/yolov7-vs-yolo26.md | 12 +++--- docs/en/compare/yolov7-vs-yolov10.md | 14 +++---- docs/en/compare/yolov7-vs-yolov5.md | 20 +++++----- docs/en/compare/yolov7-vs-yolov6.md | 24 +++++------ docs/en/compare/yolov7-vs-yolov8.md | 18 ++++----- docs/en/compare/yolov7-vs-yolov9.md | 18 ++++----- docs/en/compare/yolov7-vs-yolox.md | 18 ++++----- docs/en/compare/yolov8-vs-damo-yolo.md | 16 ++++---- docs/en/compare/yolov8-vs-efficientdet.md | 14 +++---- docs/en/compare/yolov8-vs-pp-yoloe.md | 14 +++---- docs/en/compare/yolov8-vs-rtdetr.md | 20 +++++----- docs/en/compare/yolov8-vs-yolo11.md | 22 +++++----- docs/en/compare/yolov8-vs-yolo26.md | 20 +++++----- docs/en/compare/yolov8-vs-yolov10.md | 30 +++++++------- docs/en/compare/yolov8-vs-yolov5.md | 20 +++++----- docs/en/compare/yolov8-vs-yolov6.md | 20 +++++----- docs/en/compare/yolov8-vs-yolov7.md | 6 +-- docs/en/compare/yolov8-vs-yolov9.md | 24 +++++------ docs/en/compare/yolov8-vs-yolox.md | 10 ++--- docs/en/compare/yolov9-vs-damo-yolo.md | 20 +++++----- docs/en/compare/yolov9-vs-efficientdet.md | 16 ++++---- docs/en/compare/yolov9-vs-pp-yoloe.md | 16 ++++---- docs/en/compare/yolov9-vs-rtdetr.md | 24 +++++------ docs/en/compare/yolov9-vs-yolo11.md | 22 +++++----- docs/en/compare/yolov9-vs-yolo26.md | 24 +++++------ docs/en/compare/yolov9-vs-yolov10.md | 14 +++---- docs/en/compare/yolov9-vs-yolov5.md | 20 +++++----- docs/en/compare/yolov9-vs-yolov6.md | 16 ++++---- docs/en/compare/yolov9-vs-yolov7.md | 14 +++---- docs/en/compare/yolov9-vs-yolov8.md | 18 ++++----- docs/en/compare/yolov9-vs-yolox.md | 14 +++---- docs/en/compare/yolox-vs-damo-yolo.md | 14 +++---- docs/en/compare/yolox-vs-efficientdet.md | 18 ++++----- docs/en/compare/yolox-vs-pp-yoloe.md | 20 +++++----- docs/en/compare/yolox-vs-rtdetr.md | 16 ++++---- docs/en/compare/yolox-vs-yolo11.md | 18 ++++----- docs/en/compare/yolox-vs-yolo26.md | 12 +++--- docs/en/compare/yolox-vs-yolov10.md | 12 +++--- docs/en/compare/yolox-vs-yolov5.md | 12 +++--- docs/en/compare/yolox-vs-yolov6.md | 12 +++--- docs/en/compare/yolox-vs-yolov7.md | 26 ++++++------ docs/en/compare/yolox-vs-yolov8.md | 22 +++++----- docs/en/compare/yolox-vs-yolov9.md | 18 ++++----- 160 files changed, 1447 insertions(+), 1447 deletions(-) diff --git a/.github/ISSUE_TEMPLATE/bug-report.yml b/.github/ISSUE_TEMPLATE/bug-report.yml index 2030a3b4825..967eec99b51 100644 --- a/.github/ISSUE_TEMPLATE/bug-report.yml +++ b/.github/ISSUE_TEMPLATE/bug-report.yml @@ -81,6 +81,6 @@ body: label: Are you willing to submit a PR? description: > (Optional) We encourage you to submit a [Pull Request](https://github.com/ultralytics/docs/pulls) to help improve Ultralytics Docs, especially if you know how to fix the issue. - See the Ultralytics [Contributing Guide](https://docs.ultralytics.com/help/contributing/) to get started. + See the Ultralytics [Contributing Guide](https://docs.ultralytics.com/help/contributing) to get started. options: - label: Yes I'd like to help by submitting a PR! diff --git a/.github/ISSUE_TEMPLATE/feature-request.yml b/.github/ISSUE_TEMPLATE/feature-request.yml index c077e471d14..d0e8bd8162f 100644 --- a/.github/ISSUE_TEMPLATE/feature-request.yml +++ b/.github/ISSUE_TEMPLATE/feature-request.yml @@ -46,6 +46,6 @@ body: label: Are you willing to submit a PR? description: > (Optional) We encourage you to submit a [Pull Request](https://github.com/ultralytics/docs/pulls) to help improve Ultralytics Docs. - See the Ultralytics [Contributing Guide](https://docs.ultralytics.com/help/contributing/) to get started. + See the Ultralytics [Contributing Guide](https://docs.ultralytics.com/help/contributing) to get started. options: - label: Yes I'd like to help by submitting a PR! diff --git a/README.md b/README.md index 61a29954889..5314102bce2 100644 --- a/README.md +++ b/README.md @@ -2,7 +2,7 @@ # 📚 Ultralytics Docs -Welcome to Ultralytics Docs, your comprehensive resource for understanding and utilizing our state-of-the-art [machine learning](https://www.ultralytics.com/glossary/machine-learning-ml) tools and models, including [Ultralytics YOLO](https://docs.ultralytics.com/models/yolov8/). These documents are actively maintained and deployed to [https://docs.ultralytics.com](https://docs.ultralytics.com/) for easy access. +Welcome to Ultralytics Docs, your comprehensive resource for understanding and utilizing our state-of-the-art [machine learning](https://www.ultralytics.com/glossary/machine-learning-ml) tools and models, including [Ultralytics YOLO](https://docs.ultralytics.com/models/yolov8). These documents are actively maintained and deployed to [https://docs.ultralytics.com](https://docs.ultralytics.com/) for easy access. [![Ultralytics Actions](https://github.com/ultralytics/docs/actions/workflows/format.yml/badge.svg)](https://github.com/ultralytics/docs/actions/workflows/format.yml) [![jsDelivr hits](https://data.jsdelivr.com/v1/package/gh/ultralytics/llm/badge?style=rounded)](https://www.jsdelivr.com/package/gh/ultralytics/llm) @@ -121,7 +121,7 @@ To deploy your MkDocs documentation site, choose a hosting provider and configur ## 💡 Contribute -We deeply value contributions from the open-source community to enhance Ultralytics projects. Your input helps drive innovation in [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) and [AI](https://www.ultralytics.com/glossary/artificial-intelligence-ai)! Please review our [Contributing Guide](https://docs.ultralytics.com/help/contributing/) for detailed information on how to get involved. You can also share your feedback and ideas through our quick [Survey](https://www.ultralytics.com/survey?utm_source=github&utm_medium=social&utm_campaign=Survey). A heartfelt thank you 🙏 to all our contributors for their dedication and support! +We deeply value contributions from the open-source community to enhance Ultralytics projects. Your input helps drive innovation in [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) and [AI](https://www.ultralytics.com/glossary/artificial-intelligence-ai)! Please review our [Contributing Guide](https://docs.ultralytics.com/help/contributing) for detailed information on how to get involved. You can also share your feedback and ideas through our quick [Survey](https://www.ultralytics.com/survey?utm_source=github&utm_medium=social&utm_campaign=Survey). A heartfelt thank you 🙏 to all our contributors for their dedication and support! [![Ultralytics open-source contributors](https://raw.githubusercontent.com/ultralytics/assets/main/im/image-contributors.png)](https://github.com/ultralytics/ultralytics/graphs/contributors) @@ -132,11 +132,11 @@ We look forward to your contributions! Ultralytics Docs are available under two licensing options to accommodate different usage scenarios: - **AGPL-3.0 License**: Ideal for students, researchers, and enthusiasts involved in academic pursuits and open collaboration. See the [LICENSE](https://github.com/ultralytics/docs/blob/main/LICENSE) file for full details. This license promotes sharing improvements back with the community, fostering an open and collaborative environment. -- **Enterprise License**: Designed for commercial applications, this license allows seamless integration of Ultralytics software and [AI models](https://docs.ultralytics.com/models/) into commercial products and services without the open-source requirements of AGPL-3.0. Visit [Ultralytics Licensing](https://www.ultralytics.com/license) for more information on obtaining an Enterprise License. +- **Enterprise License**: Designed for commercial applications, this license allows seamless integration of Ultralytics software and [AI models](https://docs.ultralytics.com/models) into commercial products and services without the open-source requirements of AGPL-3.0. Visit [Ultralytics Licensing](https://www.ultralytics.com/license) for more information on obtaining an Enterprise License. ## ✉️ Contact -For bug reports, feature requests, and other issues related to the documentation, please use [GitHub Issues](https://github.com/ultralytics/docs/issues). For discussions, questions, and community support regarding Ultralytics software, the [Ultralytics Platform](https://docs.ultralytics.com/platform/quickstart/), and more, join the conversation with peers and the Ultralytics team on our [Discord server](https://discord.com/invite/ultralytics)! +For bug reports, feature requests, and other issues related to the documentation, please use [GitHub Issues](https://github.com/ultralytics/docs/issues). For discussions, questions, and community support regarding Ultralytics software, the [Ultralytics Platform](https://docs.ultralytics.com/platform/quickstart), and more, join the conversation with peers and the Ultralytics team on our [Discord server](https://discord.com/invite/ultralytics)!
diff --git a/docs/en/compare/damo-yolo-vs-efficientdet.md b/docs/en/compare/damo-yolo-vs-efficientdet.md index 47311bc5a6e..b7830d090bd 100644 --- a/docs/en/compare/damo-yolo-vs-efficientdet.md +++ b/docs/en/compare/damo-yolo-vs-efficientdet.md @@ -91,17 +91,17 @@ EfficientDet is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Modern Alternative: Ultralytics YOLO26 -While both DAMO-YOLO and EfficientDet represent significant academic milestones, real-world deployment often requires a more balanced, feature-rich, and developer-friendly approach. This is where [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) sets a new industry standard. +While both DAMO-YOLO and EfficientDet represent significant academic milestones, real-world deployment often requires a more balanced, feature-rich, and developer-friendly approach. This is where [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) sets a new industry standard. -Released in January 2026, YOLO26 builds upon the legacy of its predecessors, including [Ultralytics YOLO11](https://docs.ultralytics.com/models/yolo11/) and [YOLOv8](https://docs.ultralytics.com/models/yolov8/), delivering a paradigm shift in how we approach [object detection](https://docs.ultralytics.com/tasks/detect/). +Released in January 2026, YOLO26 builds upon the legacy of its predecessors, including [Ultralytics YOLO11](https://docs.ultralytics.com/models/yolo11) and [YOLOv8](https://docs.ultralytics.com/models/yolov8), delivering a paradigm shift in how we approach [object detection](https://docs.ultralytics.com/tasks/detect). !!! tip "End-to-End Simplicity" @@ -109,13 +109,13 @@ Released in January 2026, YOLO26 builds upon the legacy of its predecessors, inc ### Unmatched Performance and Versatility -YOLO26 does not just improve on speed; it redefines training stability and accuracy. It introduces the **MuSGD Optimizer**, a hybrid of SGD and Muon inspired by LLM training innovations, leading to dramatically faster convergence rates and superior training efficiency. Unlike heavy transformer-based alternatives like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/), YOLO26 maintains incredibly low memory requirements, ensuring it can be trained on consumer-grade hardware. +YOLO26 does not just improve on speed; it redefines training stability and accuracy. It introduces the **MuSGD Optimizer**, a hybrid of SGD and Muon inspired by LLM training innovations, leading to dramatically faster convergence rates and superior training efficiency. Unlike heavy transformer-based alternatives like [RT-DETR](https://docs.ultralytics.com/models/rtdetr), YOLO26 maintains incredibly low memory requirements, ensuring it can be trained on consumer-grade hardware. Furthermore, YOLO26 incorporates **ProgLoss + STAL**, heavily improving small-object recognition which is vital for use cases like [drone aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision) and robotics. To optimize for low-power devices, YOLO26 removed the Distribution Focal Loss (DFL), resulting in up to **43% faster CPU inference** compared to previous generations. ### Ecosystem and Ease of Use -One of the largest hurdles with models like EfficientDet is the complex integration process. In contrast, the [Ultralytics Platform](https://platform.ultralytics.com) offers a well-maintained, end-to-end ecosystem. With a unified API, users can easily pivot between detection, [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +One of the largest hurdles with models like EfficientDet is the complex integration process. In contrast, the [Ultralytics Platform](https://platform.ultralytics.com) offers a well-maintained, end-to-end ecosystem. With a unified API, users can easily pivot between detection, [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb). Here is how simple it is to train and run inference with YOLO26 using the Ultralytics Python package: @@ -136,6 +136,6 @@ predictions = model.predict("image.jpg") ## Conclusion -While exploring [DAMO-YOLO vs EfficientDet](https://docs.ultralytics.com/compare/damo-yolo-vs-efficientdet/) provides excellent insights into the trade-offs between Neural Architecture Search and compound scaling, modern developers require tools that bridge the gap between academic research and production reality. +While exploring [DAMO-YOLO vs EfficientDet](https://docs.ultralytics.com/compare/damo-yolo-vs-efficientdet) provides excellent insights into the trade-offs between Neural Architecture Search and compound scaling, modern developers require tools that bridge the gap between academic research and production reality. For developers prioritizing ease of use, an active open-source community, and an uncompromised balance of speed and accuracy, **Ultralytics YOLO26** is the definitive choice. Its NMS-free architecture, low training overhead, and seamless integration with the comprehensive [Ultralytics ecosystem](https://www.ultralytics.com/) make it the ultimate framework for your next computer vision project. diff --git a/docs/en/compare/damo-yolo-vs-pp-yoloe.md b/docs/en/compare/damo-yolo-vs-pp-yoloe.md index 4c211d74057..9bd3a09569f 100644 --- a/docs/en/compare/damo-yolo-vs-pp-yoloe.md +++ b/docs/en/compare/damo-yolo-vs-pp-yoloe.md @@ -53,7 +53,7 @@ The divergence in design philosophy between these two models heavily influences ### Feature Fusion and Backbones -DAMO-YOLO's MAE-NAS generated backbones are highly tailored to edge devices, often providing a favorable speed-to-parameter ratio. However, these custom architectures can be rigid and complex to adapt for novel tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment/). The RepGFPN neck improves multi-scale feature fusion but adds complexity during the re-parameterization export phase. +DAMO-YOLO's MAE-NAS generated backbones are highly tailored to edge devices, often providing a favorable speed-to-parameter ratio. However, these custom architectures can be rigid and complex to adapt for novel tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment). The RepGFPN neck improves multi-scale feature fusion but adds complexity during the re-parameterization export phase. PP-YOLOE+ relies on the more traditional, yet highly effective, CSPRepResNet. While this backbone requires a larger parameter footprint than DAMO-YOLO for similar accuracy, it is highly stable to train and easier to integrate into existing pipelines. Its ET-head efficiently handles classification and regression, but still requires post-processing steps like Non-Maximum Suppression (NMS). @@ -84,7 +84,7 @@ As the table illustrates, DAMO-YOLO generally achieves lower latency on small (s DAMO-YOLO's reliance on distillation means you often need to train a much larger teacher model before training the smaller student model. This drastically increases the [CUDA memory requirements](https://docs.pytorch.org/docs/stable/notes/cuda.html) and overall computational budget. PP-YOLOE+ simplifies this with standard single-stage training but remains tightly coupled to the PaddlePaddle framework, which may limit flexibility for teams accustomed to PyTorch. -By contrast, the modern [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) model resolves these bottlenecks. Utilizing the new **MuSGD Optimizer**—a hybrid of SGD and Muon inspired by LLM training innovations—YOLO26 achieves faster convergence and highly stable training without requiring convoluted distillation pipelines. Additionally, YOLO models typically require far less CUDA memory during training compared to transformer-based detectors like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +By contrast, the modern [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) model resolves these bottlenecks. Utilizing the new **MuSGD Optimizer**—a hybrid of SGD and Muon inspired by LLM training innovations—YOLO26 achieves faster convergence and highly stable training without requiring convoluted distillation pipelines. Additionally, YOLO models typically require far less CUDA memory during training compared to transformer-based detectors like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). ## Real-World Applications and Ideal Use Cases @@ -104,12 +104,12 @@ When upgrading your computer vision pipeline, **Ultralytics YOLO26** provides an - **Up to 43% Faster CPU Inference:** With the complete removal of Distribution Focal Loss (DFL), YOLO26 is remarkably fast on edge CPUs and low-power IoT devices. - **Improved Small Object Detection:** The integration of ProgLoss and STAL loss functions provides dramatic improvements in small-object recognition, vital for [aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision). -- **Extensive Versatility:** Unlike PP-YOLOE+ which focuses strictly on detection, YOLO26 seamlessly handles [pose estimation](https://docs.ultralytics.com/tasks/pose/), [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), and semantic segmentation with task-specific architectural improvements. +- **Extensive Versatility:** Unlike PP-YOLOE+ which focuses strictly on detection, YOLO26 seamlessly handles [pose estimation](https://docs.ultralytics.com/tasks/pose), [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb), and semantic segmentation with task-specific architectural improvements. ## Conclusion DAMO-YOLO and PP-YOLOE+ represent important milestones in the evolution of anchor-free object detection. DAMO-YOLO pushed the limits of neural architecture search for edge latency, while PP-YOLOE+ demonstrated the power of large-scale pre-training. -However, for developers seeking the best balance of speed, accuracy, and deployment simplicity, the **Ultralytics YOLO26** model is the definitive choice. Its NMS-free architecture, robust Python API, and seamless integration with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) ensure your projects move smoothly from prototype to production. +However, for developers seeking the best balance of speed, accuracy, and deployment simplicity, the **Ultralytics YOLO26** model is the definitive choice. Its NMS-free architecture, robust Python API, and seamless integration with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) ensure your projects move smoothly from prototype to production. -Ready to get started? Explore the [Ultralytics Quickstart Guide](https://docs.ultralytics.com/quickstart/) or compare more models in our [YOLO11 vs DAMO-YOLO](https://docs.ultralytics.com/compare/yolo11-vs-damo-yolo/) overview. +Ready to get started? Explore the [Ultralytics Quickstart Guide](https://docs.ultralytics.com/quickstart) or compare more models in our [YOLO11 vs DAMO-YOLO](https://docs.ultralytics.com/compare/yolo11-vs-damo-yolo) overview. diff --git a/docs/en/compare/damo-yolo-vs-rtdetr.md b/docs/en/compare/damo-yolo-vs-rtdetr.md index d8ffd94d208..1e253829ac5 100644 --- a/docs/en/compare/damo-yolo-vs-rtdetr.md +++ b/docs/en/compare/damo-yolo-vs-rtdetr.md @@ -45,7 +45,7 @@ Baidu's RTDETRv2 represents a significant leap for Real-Time Detection Transform - **GitHub:** [RT-DETR Repository](https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetrv2_pytorch) - **Docs:** [RTDETRv2 Documentation](https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetrv2_pytorch#readme) -[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr){ .md-button } !!! tip "Embracing Transformers in Vision AI" @@ -75,22 +75,22 @@ Below is a detailed comparison of their performance metrics. Due to its NAS-optimized backbone and exceptionally low parameter count in its smaller variants (like DAMO-YOLOt), it is highly suitable for deployment on highly constrained hardware. If you are building solutions for embedded devices using runtimes like [ONNX](https://onnx.ai/) or specialized [TensorRT](https://developer.nvidia.com/tensorrt) engines for edge computing, DAMO-YOLO provides a highly responsive framework. **Where RTDETRv2 Excels:** -RTDETRv2 shines in scenarios where server-grade GPUs are available and global image context is paramount. Its transformer architecture allows it to naturally resolve overlapping bounding boxes without NMS, making it a robust choice for dense [crowd management](https://www.ultralytics.com/blog/vision-ai-in-crowd-management) or complex [object tracking](https://docs.ultralytics.com/modes/track/) where spatial relationships between distant objects are critical. +RTDETRv2 shines in scenarios where server-grade GPUs are available and global image context is paramount. Its transformer architecture allows it to naturally resolve overlapping bounding boxes without NMS, making it a robust choice for dense [crowd management](https://www.ultralytics.com/blog/vision-ai-in-crowd-management) or complex [object tracking](https://docs.ultralytics.com/modes/track) where spatial relationships between distant objects are critical. ## The Ultralytics Advantage: Introducing YOLO26 While DAMO-YOLO and RTDETRv2 represent significant academic achievements, transitioning these models into scalable, production-ready applications can be challenging. Developers often face fragmented codebases, lack of support for multi-task learning, and complicated deployment pipelines. -This is where the [Ultralytics ecosystem](https://docs.ultralytics.com/platform/) truly sets itself apart. By prioritizing ease of use, a well-maintained Python API, and unmatched versatility, Ultralytics ensures that developers spend less time debugging and more time building. +This is where the [Ultralytics ecosystem](https://docs.ultralytics.com/platform) truly sets itself apart. By prioritizing ease of use, a well-maintained Python API, and unmatched versatility, Ultralytics ensures that developers spend less time debugging and more time building. The recently released **Ultralytics YOLO26** model takes these advantages to the next level, offering breakthroughs that outpace both DAMO-YOLO and RTDETRv2: -- **End-to-End NMS-Free Design:** Pioneered originally in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 is natively end-to-end. This completely eliminates NMS post-processing, making deployment faster and drastically simpler than traditional CNNs, while matching the direct-output benefits of RTDETRv2. +- **End-to-End NMS-Free Design:** Pioneered originally in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 is natively end-to-end. This completely eliminates NMS post-processing, making deployment faster and drastically simpler than traditional CNNs, while matching the direct-output benefits of RTDETRv2. - **Up to 43% Faster CPU Inference:** Optimized heavily for [edge AI devices](https://www.ultralytics.com/blog/edge-ai-and-edge-computing-powering-real-time-intelligence) without discrete GPUs, making it a vastly superior choice for IoT applications compared to memory-heavy transformers. - **MuSGD Optimizer:** Inspired by Moonshot AI's Kimi K2, this hybrid of SGD and Muon brings Large Language Model (LLM) training innovations into computer vision, resulting in remarkably stable training and faster convergence. - **ProgLoss + STAL:** These advanced loss functions deliver notable improvements in small-object recognition, an area where models traditionally struggle. This is critical for [aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision) and drone applications. - **DFL Removal:** Distribution Focal Loss has been removed to ensure simplified export formats and better compatibility with low-power edge devices. -- **Unrivaled Versatility:** Unlike competing models limited strictly to detection, YOLO26 includes task-specific improvements across the board, such as specialized angle loss for [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), semantic segmentation loss for pixel-perfect accuracy, and Residual Log-Likelihood Estimation (RLE) for [Pose estimation](https://docs.ultralytics.com/tasks/pose/). +- **Unrivaled Versatility:** Unlike competing models limited strictly to detection, YOLO26 includes task-specific improvements across the board, such as specialized angle loss for [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb), semantic segmentation loss for pixel-perfect accuracy, and Residual Log-Likelihood Estimation (RLE) for [Pose estimation](https://docs.ultralytics.com/tasks/pose). [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -119,7 +119,7 @@ results_yolo = model_yolo("https://ultralytics.com/images/bus.jpg") results_yolo[0].show() ``` -This simplicity extends to [custom dataset training](https://docs.ultralytics.com/guides/custom-trainer/) and exporting. Utilizing the [Ultralytics Python package](https://docs.ultralytics.com/usage/python/), developers can easily push their trained weights to deployment platforms like [CoreML](https://docs.ultralytics.com/integrations/coreml/) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) with a single command. +This simplicity extends to [custom dataset training](https://docs.ultralytics.com/guides/custom-trainer) and exporting. Utilizing the [Ultralytics Python package](https://docs.ultralytics.com/usage/python), developers can easily push their trained weights to deployment platforms like [CoreML](https://docs.ultralytics.com/integrations/coreml) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino) with a single command. ## Conclusion and Further Exploration @@ -127,4 +127,4 @@ Both DAMO-YOLO and RTDETRv2 have undeniably pushed the boundaries of what is pos However, for developers seeking the ultimate balance of performance, comprehensive documentation, and production readiness, **Ultralytics YOLO models** remain the gold standard. With the introduction of YOLO26, users gain access to transformer-like end-to-end detection, LLM-inspired training efficiency, and unparalleled CPU speeds—all wrapped within an intuitive and robust ecosystem. -If you are evaluating models for your next project, you may also find value in reading our comparisons of [EfficientDet vs RTDETR](https://docs.ultralytics.com/compare/efficientdet-vs-rtdetr/), exploring the previous generation [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), or reviewing academic baselines like [YOLOX](https://docs.ultralytics.com/compare/yolox-vs-rtdetr/). Start building today by exploring the [Ultralytics quickstart guide](https://docs.ultralytics.com/quickstart/). +If you are evaluating models for your next project, you may also find value in reading our comparisons of [EfficientDet vs RTDETR](https://docs.ultralytics.com/compare/efficientdet-vs-rtdetr), exploring the previous generation [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), or reviewing academic baselines like [YOLOX](https://docs.ultralytics.com/compare/yolox-vs-rtdetr). Start building today by exploring the [Ultralytics quickstart guide](https://docs.ultralytics.com/quickstart). diff --git a/docs/en/compare/damo-yolo-vs-yolo11.md b/docs/en/compare/damo-yolo-vs-yolo11.md index 07b19d1b742..41cbe148729 100644 --- a/docs/en/compare/damo-yolo-vs-yolo11.md +++ b/docs/en/compare/damo-yolo-vs-yolo11.md @@ -21,7 +21,7 @@ Authors: Glenn Jocher and Jing Qiu Organization: [Ultralytics](https://www.ultralytics.com) Date: 2024-09-27 GitHub: [ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -Docs: [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11/) +Docs: [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11) @@ -34,7 +34,7 @@ The underlying architecture of an object detection model dictates its inference **DAMO-YOLO** introduces several academic innovations, heavily relying on Neural Architecture Search (NAS) to automatically design its backbone. It utilizes an efficient RepGFPN (Reparameterized Generalized Feature Pyramid Network) to enhance feature fusion and a ZeroHead design that significantly scales down the heavy prediction head often found in previous architectures. While this NAS-driven approach allows DAMO-YOLO to achieve specific efficiencies on selected GPUs, the resulting architectures can sometimes lack the flexibility needed to generalize seamlessly across diverse edge devices. -In contrast, **YOLO11** builds upon years of foundational research to deliver a highly optimized, handcrafted architecture. It focuses on a streamlined backbone and a highly efficient neck that reduces redundant computations. One of the primary advantages of YOLO11 is its refined parameter efficiency; it achieves high feature representation without the heavy VRAM requirements typical of transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). This makes YOLO11 exceptionally versatile, capable of running smoothly on consumer-grade GPUs, mobile devices, and specialized edge accelerators. +In contrast, **YOLO11** builds upon years of foundational research to deliver a highly optimized, handcrafted architecture. It focuses on a streamlined backbone and a highly efficient neck that reduces redundant computations. One of the primary advantages of YOLO11 is its refined parameter efficiency; it achieves high feature representation without the heavy VRAM requirements typical of transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). This makes YOLO11 exceptionally versatile, capable of running smoothly on consumer-grade GPUs, mobile devices, and specialized edge accelerators. ## Performance and Metrics @@ -63,7 +63,7 @@ The training pipeline is where developers spend the majority of their time, maki DAMO-YOLO employs a multi-stage training process heavily dependent on knowledge distillation. It utilizes AlignedOTA (Optimal Transport Assignment) for label assignment and often requires training a larger "teacher" model to distill knowledge into the smaller "student" models. This methodology drastically increases the [CUDA memory](https://developer.nvidia.com/cuda) footprint and the overall compute time required to achieve optimal convergence. -Conversely, the Ultralytics ecosystem abstracts away the complexity of model training. YOLO11 is designed for exceptional ease of use, featuring a streamlined Python API and comprehensive [CLI interfaces](https://docs.ultralytics.com/usage/cli/) that allow engineers to initiate training on custom datasets with a single command. The training pipeline is inherently resource-efficient, minimizing memory spikes so that even larger models can be trained on standard hardware. +Conversely, the Ultralytics ecosystem abstracts away the complexity of model training. YOLO11 is designed for exceptional ease of use, featuring a streamlined Python API and comprehensive [CLI interfaces](https://docs.ultralytics.com/usage/cli) that allow engineers to initiate training on custom datasets with a single command. The training pipeline is inherently resource-efficient, minimizing memory spikes so that even larger models can be trained on standard hardware. !!! tip "Streamlined Training with Ultralytics" @@ -96,13 +96,13 @@ DAMO-YOLO is strictly an object detection framework. It excels in academic resea ### The Ultralytics Advantage -Ultralytics models, including YOLO11, shine in real-world commercial applications due to their unparalleled versatility and well-maintained ecosystem. Unlike DAMO-YOLO, the Ultralytics framework supports multi-modal tasks natively. From [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/) in medical imaging to [Pose Estimation](https://docs.ultralytics.com/tasks/pose/) for biomechanical analysis in sports, a single, unified codebase handles it all. +Ultralytics models, including YOLO11, shine in real-world commercial applications due to their unparalleled versatility and well-maintained ecosystem. Unlike DAMO-YOLO, the Ultralytics framework supports multi-modal tasks natively. From [Instance Segmentation](https://docs.ultralytics.com/tasks/segment) in medical imaging to [Pose Estimation](https://docs.ultralytics.com/tasks/pose) for biomechanical analysis in sports, a single, unified codebase handles it all. Industries leveraging YOLO11 include: - **Smart Agriculture:** Utilizing object detection to monitor crop health and automate harvesting machinery. - **Retail Analytics:** Implementing [smart surveillance](https://www.ultralytics.com/blog/smart-surveillance-ultralytics-yolo11) to analyze customer traffic and automate inventory management. -- **Logistics and Supply Chain:** High-speed barcode and package detection using [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) on fast-moving conveyor belts. +- **Logistics and Supply Chain:** High-speed barcode and package detection using [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb) on fast-moving conveyor belts. ## Use Cases and Recommendations @@ -120,17 +120,17 @@ DAMO-YOLO is a strong choice for: YOLO11 is recommended for: -- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) where reliability and active maintenance are paramount. -- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [OBB](https://docs.ultralytics.com/tasks/obb/) within a single unified framework. -- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python/). +- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) where reliability and active maintenance are paramount. +- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [OBB](https://docs.ultralytics.com/tasks/obb) within a single unified framework. +- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python). ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Next Generation: Introducing YOLO26 @@ -147,6 +147,6 @@ Released in January 2026, YOLO26 introduces several groundbreaking advancements: ## Conclusion -Both DAMO-YOLO and YOLO11 have contributed significantly to the advancement of fast, accurate computer vision. While DAMO-YOLO offers interesting academic insights into architecture search and distillation, Ultralytics YOLO11 (and the groundbreaking [YOLO26](https://docs.ultralytics.com/models/yolo26/)) provides a superior developer experience. +Both DAMO-YOLO and YOLO11 have contributed significantly to the advancement of fast, accurate computer vision. While DAMO-YOLO offers interesting academic insights into architecture search and distillation, Ultralytics YOLO11 (and the groundbreaking [YOLO26](https://docs.ultralytics.com/models/yolo26)) provides a superior developer experience. -With lower memory requirements, extensive documentation, multi-task capabilities, and integration with the powerful [Ultralytics Platform](https://platform.ultralytics.com), Ultralytics models remain the top recommendation for researchers and enterprise engineers looking to build robust, scalable AI solutions. For those exploring other advanced architectures, comparing [YOLO26 vs RT-DETR](https://docs.ultralytics.com/compare/yolo26-vs-rtdetr/) offers additional insights into transformer-based alternatives. +With lower memory requirements, extensive documentation, multi-task capabilities, and integration with the powerful [Ultralytics Platform](https://platform.ultralytics.com), Ultralytics models remain the top recommendation for researchers and enterprise engineers looking to build robust, scalable AI solutions. For those exploring other advanced architectures, comparing [YOLO26 vs RT-DETR](https://docs.ultralytics.com/compare/yolo26-vs-rtdetr) offers additional insights into transformer-based alternatives. diff --git a/docs/en/compare/damo-yolo-vs-yolo26.md b/docs/en/compare/damo-yolo-vs-yolo26.md index a804e4f7d31..a001f45ac08 100644 --- a/docs/en/compare/damo-yolo-vs-yolo26.md +++ b/docs/en/compare/damo-yolo-vs-yolo26.md @@ -34,27 +34,27 @@ DAMO-YOLO introduces several technical innovations designed to push the boundari ## The Ultralytics Advantage: YOLO26 -Released on January 14, 2026, by Glenn Jocher and Jing Qiu at [Ultralytics](https://www.ultralytics.com/), **YOLO26** represents the pinnacle of accessible, high-performance vision AI. Building upon the legacy of [YOLO11](https://docs.ultralytics.com/models/yolo11/) and [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 is designed from the ground up for edge-first deployment, multimodal versatility, and unparalleled ease of use. +Released on January 14, 2026, by Glenn Jocher and Jing Qiu at [Ultralytics](https://www.ultralytics.com/), **YOLO26** represents the pinnacle of accessible, high-performance vision AI. Building upon the legacy of [YOLO11](https://docs.ultralytics.com/models/yolo11) and [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 is designed from the ground up for edge-first deployment, multimodal versatility, and unparalleled ease of use. ### YOLO26 Innovations Ultralytics YOLO26 introduces several groundbreaking features that make it the definitive choice for modern computer vision applications: - **End-to-End NMS-Free Design:** YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing. Pioneered initially in YOLOv10, this end-to-end approach drastically simplifies deployment pipelines and ensures deterministic, low-latency inference. -- **Up to 43% Faster CPU Inference:** Architecturally optimized for edge computing, YOLO26 delivers exceptional speed on edge devices and standard [CPUs](https://docs.ultralytics.com/reference/utils/cpu/), making it perfect for battery-powered IoT devices. +- **Up to 43% Faster CPU Inference:** Architecturally optimized for edge computing, YOLO26 delivers exceptional speed on edge devices and standard [CPUs](https://docs.ultralytics.com/reference/utils/cpu), making it perfect for battery-powered IoT devices. - **MuSGD Optimizer:** Inspired by LLM training (like Moonshot AI's Kimi K2), YOLO26 incorporates a hybrid of SGD and Muon. This brings large language model training stability to computer vision, resulting in faster and more reliable convergence. -- **DFL Removal:** By removing Distribution Focal Loss, the model graph is simplified, allowing for frictionless export to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). +- **DFL Removal:** By removing Distribution Focal Loss, the model graph is simplified, allowing for frictionless export to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). - **ProgLoss + STAL:** These advanced loss functions provide notable improvements in small-object recognition, a critical feature for [drone operations](https://www.ultralytics.com/blog/computer-vision-applications-ai-drone-uav-operations) and [agriculture](https://www.ultralytics.com/solutions/ai-in-agriculture). !!! tip "Task-Specific Enhancements" - YOLO26 includes specialized improvements across multiple modalities: a multi-scale proto for [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), Residual Log-Likelihood Estimation (RLE) for [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), and advanced angle loss to mitigate boundary issues in [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. + YOLO26 includes specialized improvements across multiple modalities: a multi-scale proto for [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), Residual Log-Likelihood Estimation (RLE) for [Pose Estimation](https://docs.ultralytics.com/tasks/pose), and advanced angle loss to mitigate boundary issues in [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } ## Performance Comparison -When evaluating these models, the balance between accuracy (mAP) and computational efficiency (Speed/FLOPs) is paramount. The table below highlights how these models compare using the industry-standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). +When evaluating these models, the balance between accuracy (mAP) and computational efficiency (Speed/FLOPs) is paramount. The table below highlights how these models compare using the industry-standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -79,7 +79,7 @@ While DAMO-YOLO achieves competitive accuracy, its training methodology is highl ### The Streamlined Ultralytics Experience -Conversely, Ultralytics YOLO26 is designed for "zero-to-hero" usability. The entire training, validation, and deployment lifecycle is abstracted behind a clean, unified Python API and CLI. Furthermore, YOLO26 requires significantly less [CUDA](https://developer.nvidia.com/cuda/toolkit) memory during training compared to transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/), allowing researchers to train state-of-the-art models on consumer-grade hardware. +Conversely, Ultralytics YOLO26 is designed for "zero-to-hero" usability. The entire training, validation, and deployment lifecycle is abstracted behind a clean, unified Python API and CLI. Furthermore, YOLO26 requires significantly less [CUDA](https://developer.nvidia.com/cuda/toolkit) memory during training compared to transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr), allowing researchers to train state-of-the-art models on consumer-grade hardware. Here is an example of how simple it is to train, evaluate, and export a YOLO26 model using the Ultralytics SDK: @@ -115,7 +115,7 @@ For high-speed [manufacturing automation](https://www.ultralytics.com/blog/manuf ### Edge AI and Mobile Devices -Deploying computer vision on battery-powered devices requires extreme efficiency. While DAMO-YOLO relies on specific RepGFPN necks, **YOLO26n** (Nano) is specifically optimized for edge computing. Its DFL removal and **43% faster CPU inference** make it the ultimate solution for smart cameras, mobile applications, and [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/). +Deploying computer vision on battery-powered devices requires extreme efficiency. While DAMO-YOLO relies on specific RepGFPN necks, **YOLO26n** (Nano) is specifically optimized for edge computing. Its DFL removal and **43% faster CPU inference** make it the ultimate solution for smart cameras, mobile applications, and [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system). ### Multi-Modal Project Requirements @@ -139,7 +139,7 @@ YOLO26 is recommended for: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Conclusion @@ -147,4 +147,4 @@ Both architectures represent significant achievements in the field of deep learn However, for developers, researchers, and enterprises looking for a production-ready solution, **Ultralytics YOLO26** stands out as the superior choice. Its combination of an end-to-end NMS-free design, massive CPU inference gains, multimodal versatility, and integration into the well-maintained Ultralytics ecosystem makes it the most robust and practical tool for solving real-world computer vision challenges today. -For users interested in exploring other models within the Ultralytics ecosystem, comprehensive documentation is available for [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), and the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +For users interested in exploring other models within the Ultralytics ecosystem, comprehensive documentation is available for [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), and the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr). diff --git a/docs/en/compare/damo-yolo-vs-yolov10.md b/docs/en/compare/damo-yolo-vs-yolov10.md index 67c44339436..8906426424e 100644 --- a/docs/en/compare/damo-yolo-vs-yolov10.md +++ b/docs/en/compare/damo-yolo-vs-yolov10.md @@ -6,7 +6,7 @@ keywords: YOLOv10, DAMO-YOLO, object detection comparison, YOLO models, DAMO-YOL # DAMO-YOLO vs YOLOv10: Evolution of Efficient Real-Time Object Detection -The field of computer vision has witnessed a rapid evolution in real-time [object detection](https://docs.ultralytics.com/tasks/detect/) architectures. When comparing **DAMO-YOLO** and **YOLOv10**, we observe two distinct philosophies in model design: automated architecture search versus end-to-end NMS-free optimization. While both push the boundaries of accuracy and speed, their underlying structures and ideal use cases differ significantly. +The field of computer vision has witnessed a rapid evolution in real-time [object detection](https://docs.ultralytics.com/tasks/detect) architectures. When comparing **DAMO-YOLO** and **YOLOv10**, we observe two distinct philosophies in model design: automated architecture search versus end-to-end NMS-free optimization. While both push the boundaries of accuracy and speed, their underlying structures and ideal use cases differ significantly. @@ -42,19 +42,19 @@ Released a year and a half later, YOLOv10 introduced a paradigm shift by complet - **Organization:** [Tsinghua University](https://www.tsinghua.edu.cn/en/) - **Date:** May 23, 2024 - **Arxiv:** [2405.14458](https://arxiv.org/abs/2405.14458) -- **Docs:** [Ultralytics YOLOv10](https://docs.ultralytics.com/models/yolov10/) +- **Docs:** [Ultralytics YOLOv10](https://docs.ultralytics.com/models/yolov10) ### Architectural Highlights The standout feature of YOLOv10 is its **consistent dual assignments** for NMS-free training. Traditional detectors predict multiple overlapping bounding boxes for a single object, requiring NMS to filter duplicates. This post-processing step creates a bottleneck, especially on edge devices. YOLOv10 solves this by allowing the model to naturally predict a single, accurate bounding box per object. -The authors also focused on a holistic efficiency-accuracy driven model design. By carefully analyzing the computational redundancy in existing architectures, they optimized the backbone and head to reduce the number of [FLOPs](https://www.ultralytics.com/glossary/flops) and parameters. This lightweight design ensures YOLOv10 delivers exceptional inference latency when exported to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino/). +The authors also focused on a holistic efficiency-accuracy driven model design. By carefully analyzing the computational redundancy in existing architectures, they optimized the backbone and head to reduce the number of [FLOPs](https://www.ultralytics.com/glossary/flops) and parameters. This lightweight design ensures YOLOv10 delivers exceptional inference latency when exported to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino). -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## Performance and Benchmarks -The table below illustrates the raw performance metrics on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). Best overall values in each column are highlighted in **bold**. +The table below illustrates the raw performance metrics on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). Best overall values in each column are highlighted in **bold**. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -70,13 +70,13 @@ The table below illustrates the raw performance metrics on the [COCO dataset](ht | YOLOv10l | 640 | 53.3 | - | 8.33 | 29.5 | 120.3 | | YOLOv10x | 640 | **54.4** | - | 12.2 | 56.9 | 160.4 | -While DAMO-YOLO holds its own in terms of accuracy, YOLOv10 consistently provides lower latency and significantly smaller [model weights](https://www.ultralytics.com/glossary/model-weights). For instance, YOLOv10s achieves a slightly higher mAP (46.7%) than DAMO-YOLOs (46.0%) while using fewer than half the parameters (7.2M vs 16.3M). The lower [memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics/) make YOLOv10 an exceptionally versatile choice for embedded systems. +While DAMO-YOLO holds its own in terms of accuracy, YOLOv10 consistently provides lower latency and significantly smaller [model weights](https://www.ultralytics.com/glossary/model-weights). For instance, YOLOv10s achieves a slightly higher mAP (46.7%) than DAMO-YOLOs (46.0%) while using fewer than half the parameters (7.2M vs 16.3M). The lower [memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics) make YOLOv10 an exceptionally versatile choice for embedded systems. ## Training Efficiency and Usability When transitioning from academic research to production, ease of use is paramount. DAMO-YOLO's multi-stage distillation process and complex NAS configurations can pose steep learning curves for engineering teams. -Conversely, YOLOv10 benefits immensely from being fully integrated into the [Ultralytics Python SDK](https://docs.ultralytics.com/usage/python/). Training a custom model involves minimal boilerplate code. Ultralytics handles [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation/), [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/), and [experiment tracking](https://www.ultralytics.com/glossary/experiment-tracking) automatically. +Conversely, YOLOv10 benefits immensely from being fully integrated into the [Ultralytics Python SDK](https://docs.ultralytics.com/usage/python). Training a custom model involves minimal boilerplate code. Ultralytics handles [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation), [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning), and [experiment tracking](https://www.ultralytics.com/glossary/experiment-tracking) automatically. ```python from ultralytics import YOLO @@ -94,7 +94,7 @@ prediction[0].show() !!! tip "Fast Prototyping" - Using the Ultralytics ecosystem allows developers to move from a prototype to a fully [exported ONNX model](https://docs.ultralytics.com/integrations/onnx/) in just a few lines of code, bypassing the complex environment setups required by older frameworks. + Using the Ultralytics ecosystem allows developers to move from a prototype to a fully [exported ONNX model](https://docs.ultralytics.com/integrations/onnx) in just a few lines of code, bypassing the complex environment setups required by older frameworks. ## Real-World Use Cases @@ -124,11 +124,11 @@ YOLOv10 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Next Generation: Enter Ultralytics YOLO26 @@ -138,8 +138,8 @@ YOLO26 features a strictly natively end-to-end design, eliminating NMS post-proc On the training side, YOLO26 introduces the **MuSGD Optimizer**, a hybrid inspired by Large Language Model (LLM) training techniques. This ensures more stable training and faster convergence. Coupled with the **ProgLoss + STAL** loss functions, YOLO26 exhibits remarkable improvements in small-object recognition, a critical feature for [wildlife conservation](https://www.ultralytics.com/blog/ai-in-wildlife-conservation) and [drone operations](https://www.ultralytics.com/blog/computer-vision-applications-ai-drone-uav-operations). -Crucially, YOLO26 is not just an object detector. It offers task-specific improvements across the board, natively supporting [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/) using Residual Log-Likelihood Estimation (RLE), and specialized angle losses for [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). With up to 43% faster CPU inference than its predecessors, it is the definitive choice for agile engineering teams. +Crucially, YOLO26 is not just an object detector. It offers task-specific improvements across the board, natively supporting [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose) using Residual Log-Likelihood Estimation (RLE), and specialized angle losses for [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). With up to 43% faster CPU inference than its predecessors, it is the definitive choice for agile engineering teams. For centralized management, annotation, and cloud training of YOLO26 models, the [Ultralytics Platform](https://platform.ultralytics.com/) provides an intuitive interface that streamlines the entire computer vision lifecycle. -Developers interested in exploring other recent advancements can also evaluate [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) framework for scenarios requiring distinct architectural solutions. +Developers interested in exploring other recent advancements can also evaluate [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr) framework for scenarios requiring distinct architectural solutions. diff --git a/docs/en/compare/damo-yolo-vs-yolov5.md b/docs/en/compare/damo-yolo-vs-yolov5.md index 60e093947f1..b48d7035fe8 100644 --- a/docs/en/compare/damo-yolo-vs-yolov5.md +++ b/docs/en/compare/damo-yolo-vs-yolov5.md @@ -40,7 +40,7 @@ Ultralytics YOLOv5 is one of the most widely adopted vision architectures in the - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** June 26, 2020 - **GitHub:** [ultralytics/yolov5](https://github.com/ultralytics/yolov5) -- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5/) +- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5) [Learn more about YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5){ .md-button } @@ -48,7 +48,7 @@ Ultralytics YOLOv5 is one of the most widely adopted vision architectures in the YOLOv5 redefined the industry standard for usability. Built natively in [PyTorch](https://pytorch.org/), it utilizes a highly optimized CSPNet backbone and a PANet neck for robust feature aggregation. While it preceded the anchor-free trend seen in later models, its highly refined anchor-based approach, coupled with automatic anchor learning, ensures excellent performance out of the box. -The true strength of YOLOv5 lies in its **Well-Maintained Ecosystem**. It seamlessly integrates with tracking tools like [Comet](https://docs.ultralytics.com/integrations/comet/) and [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/), and supports one-click exports to formats such as [ONNX](https://onnx.ai/), [TensorRT](https://developer.nvidia.com/tensorrt), and [CoreML](https://docs.ultralytics.com/integrations/coreml/). +The true strength of YOLOv5 lies in its **Well-Maintained Ecosystem**. It seamlessly integrates with tracking tools like [Comet](https://docs.ultralytics.com/integrations/comet) and [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases), and supports one-click exports to formats such as [ONNX](https://onnx.ai/), [TensorRT](https://developer.nvidia.com/tensorrt), and [CoreML](https://docs.ultralytics.com/integrations/coreml). !!! tip "Getting Started with YOLOv5" @@ -75,7 +75,7 @@ When comparing these models, it is crucial to look at the balance of mean Averag DAMO-YOLO achieves impressive mAP scores for its parameter sizes, heavily benefiting from its distillation training phase. However, this comes at the cost of **Training Efficiency**. The multi-stage distillation process requires training a heavy teacher model first, which significantly increases the necessary [GPU compute](https://www.ultralytics.com/glossary/gpu-graphics-processing-unit) time and VRAM. -Conversely, **YOLOv5** offers excellent **Memory Requirements**. Ultralytics YOLO models are known for lower memory usage during both training and inference compared to complex distillation pipelines or transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). This allows YOLOv5 to be trained efficiently on consumer-grade hardware or accessible cloud environments like [Google Colab](https://colab.research.google.com/). +Conversely, **YOLOv5** offers excellent **Memory Requirements**. Ultralytics YOLO models are known for lower memory usage during both training and inference compared to complex distillation pipelines or transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). This allows YOLOv5 to be trained efficiently on consumer-grade hardware or accessible cloud environments like [Google Colab](https://colab.research.google.com/). ## Real-World Applications and Versatility @@ -83,13 +83,13 @@ Choosing the right architecture often depends on the deployment environment. ### Where DAMO-YOLO Excels -DAMO-YOLO is strictly an [object detection](https://docs.ultralytics.com/tasks/detect/) model. It is an excellent choice for academic research, particularly for teams studying Neural Architecture Search or those aiming to reproduce the rep-parameterization techniques detailed in the paper. If a project has extensive computational resources to execute the distillation training phase and is focused solely on squeezing out the last fraction of accuracy for 2D bounding boxes, DAMO-YOLO is a strong contender. +DAMO-YOLO is strictly an [object detection](https://docs.ultralytics.com/tasks/detect) model. It is an excellent choice for academic research, particularly for teams studying Neural Architecture Search or those aiming to reproduce the rep-parameterization techniques detailed in the paper. If a project has extensive computational resources to execute the distillation training phase and is focused solely on squeezing out the last fraction of accuracy for 2D bounding boxes, DAMO-YOLO is a strong contender. ### The Ultralytics Advantage -For real-world production, the **Ease of Use** and **Versatility** of Ultralytics models make them the preferred choice. While YOLOv5 remains a staple for detection and [image classification](https://docs.ultralytics.com/tasks/classify/), the broader Ultralytics ecosystem allows developers to effortlessly switch between tasks. +For real-world production, the **Ease of Use** and **Versatility** of Ultralytics models make them the preferred choice. While YOLOv5 remains a staple for detection and [image classification](https://docs.ultralytics.com/tasks/classify), the broader Ultralytics ecosystem allows developers to effortlessly switch between tasks. -For instance, newer iterations in the Ultralytics family natively support [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. This multi-task capability ensures that teams can utilize a single, unified Python API for complex pipelines, such as combining [automated number plate recognition](https://www.ultralytics.com/blog/using-ultralytics-yolo11-for-automatic-number-plate-recognition) with vehicle segmentation. +For instance, newer iterations in the Ultralytics family natively support [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. This multi-task capability ensures that teams can utilize a single, unified Python API for complex pipelines, such as combining [automated number plate recognition](https://www.ultralytics.com/blog/using-ultralytics-yolo11-for-automatic-number-plate-recognition) with vehicle segmentation. ## Use Cases and Recommendations @@ -109,15 +109,15 @@ YOLOv5 is recommended for: - **Proven Production Systems:** Existing deployments where YOLOv5's long track record of stability, extensive documentation, and massive community support are valued. - **Resource-Constrained Training:** Environments with limited GPU resources where YOLOv5's efficient training pipeline and lower memory requirements are advantageous. -- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), [CoreML](https://docs.ultralytics.com/integrations/coreml/), and [TFLite](https://docs.ultralytics.com/integrations/tflite/). +- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), [CoreML](https://docs.ultralytics.com/integrations/coreml), and [TFLite](https://docs.ultralytics.com/integrations/tflite). ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Future: Moving to YOLO26 @@ -130,7 +130,7 @@ YOLO26 addresses the traditional bottlenecks of edge deployment and training ins - **End-to-End NMS-Free Design:** YOLO26 natively eliminates Non-Maximum Suppression post-processing. This breakthrough simplifies deployment logic and drastically reduces latency variability, making it ideal for high-speed [robotics](https://www.ultralytics.com/glossary/robotics) and autonomous systems. - **MuSGD Optimizer:** Inspired by LLM training innovations (like Moonshot AI's Kimi K2), YOLO26 utilizes the MuSGD optimizer (a hybrid of SGD and Muon). This ensures highly stable training runs and remarkably faster convergence. - **Up to 43% Faster CPU Inference:** By strategically removing the Distribution Focal Loss (DFL), YOLO26 achieves vastly superior speeds on CPUs and edge devices compared to its predecessors like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8). -- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, which is critical for analyzing [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and IoT sensor feeds. +- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, which is critical for analyzing [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and IoT sensor feeds. ### Code Example: Simplicity in Action @@ -157,4 +157,4 @@ model.export(format="onnx") Both DAMO-YOLO and YOLOv5 have contributed significantly to the landscape of computer vision. DAMO-YOLO showcases the power of Neural Architecture Search and distillation, making it an interesting study for researchers. However, **YOLOv5** remains a practical powerhouse due to its **Performance Balance**, low memory requirements, and unmatched ease of use. -For developers starting new projects today, the recommendation is to leverage the [Ultralytics Platform](https://platform.ultralytics.com) and adopt **YOLO26**. It combines the beloved user-friendly ecosystem of YOLOv5 with groundbreaking architectural advancements, ensuring top-tier accuracy and blazing-fast inference for both cloud and edge AI applications. Developers may also want to explore other efficient models like [YOLOv6](https://docs.ultralytics.com/models/yolov6/) or [YOLOX](https://docs.ultralytics.com/) depending on specific legacy hardware constraints. +For developers starting new projects today, the recommendation is to leverage the [Ultralytics Platform](https://platform.ultralytics.com) and adopt **YOLO26**. It combines the beloved user-friendly ecosystem of YOLOv5 with groundbreaking architectural advancements, ensuring top-tier accuracy and blazing-fast inference for both cloud and edge AI applications. Developers may also want to explore other efficient models like [YOLOv6](https://docs.ultralytics.com/models/yolov6) or [YOLOX](https://docs.ultralytics.com/) depending on specific legacy hardware constraints. diff --git a/docs/en/compare/damo-yolo-vs-yolov6.md b/docs/en/compare/damo-yolo-vs-yolov6.md index 250fec29b92..c50c4dd447b 100644 --- a/docs/en/compare/damo-yolo-vs-yolov6.md +++ b/docs/en/compare/damo-yolo-vs-yolov6.md @@ -37,7 +37,7 @@ Furthermore, the model introduces a "ZeroHead" design. By removing complex multi ## YOLOv6-3.0: Maximizing Industrial Throughput -Pioneered by the Meituan Vision AI Department, [YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6/) is explicitly labeled as an industrial object detector, engineered specifically to maximize throughput on NVIDIA hardware. +Pioneered by the Meituan Vision AI Department, [YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6) is explicitly labeled as an industrial object detector, engineered specifically to maximize throughput on NVIDIA hardware. - **Authors:** Chuyi Li, Lulu Li, Yifei Geng, Hongliang Jiang, Meng Cheng, Bo Zhang, Zaidan Ke, Xiaoming Xu, and Xiangxiang Chu - **Organization:** [Meituan](https://www.meituan.com/) @@ -47,11 +47,11 @@ Pioneered by the Meituan Vision AI Department, [YOLOv6-3.0](https://docs.ultraly ### Key Features and Enhancements -YOLOv6-3.0 is built upon the hardware-friendly EfficientRep backbone, making it exceptionally fast when leveraging optimizations like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) on modern GPUs. In its v3.0 iteration, the network integrates a Bi-directional Concatenation (BiC) module to improve the localization of varying object sizes. +YOLOv6-3.0 is built upon the hardware-friendly EfficientRep backbone, making it exceptionally fast when leveraging optimizations like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) on modern GPUs. In its v3.0 iteration, the network integrates a Bi-directional Concatenation (BiC) module to improve the localization of varying object sizes. Another standout feature is the Anchor-Aided Training (AAT) strategy. AAT combines the stability of [anchor-based detectors](https://www.ultralytics.com/glossary/anchor-based-detectors) during training with the inference speed of an anchor-free design. This hybrid approach yields excellent convergence without sacrificing deployment latency, making it a powerful choice for processing massive video streams in smart city analytics and automated checkout systems. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Performance Comparison @@ -97,33 +97,33 @@ YOLOv6 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Introducing YOLO26 -While both DAMO-YOLO and YOLOv6-3.0 are highly capable, they suffer from fragmented ecosystems, single-task limitations, and complex deployment pipelines. For modern engineering teams, [Ultralytics models](https://docs.ultralytics.com/models/) provide a substantially better developer experience, culminating in the groundbreaking **YOLO26**. +While both DAMO-YOLO and YOLOv6-3.0 are highly capable, they suffer from fragmented ecosystems, single-task limitations, and complex deployment pipelines. For modern engineering teams, [Ultralytics models](https://docs.ultralytics.com/models) provide a substantially better developer experience, culminating in the groundbreaking **YOLO26**. -Released in January 2026, [YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) represents the new standard for edge and cloud deployment, heavily optimizing [memory requirements](https://docs.ultralytics.com/guides/model-training-tips/) and computational efficiency. +Released in January 2026, [YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) represents the new standard for edge and cloud deployment, heavily optimizing [memory requirements](https://docs.ultralytics.com/guides/model-training-tips) and computational efficiency. ### Why Choose YOLO26? -1. **End-to-End NMS-Free Design:** Building on concepts from [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates Non-Maximum Suppression post-processing. This significantly simplifies deployment code and reduces inference latency variance across all edge devices. +1. **End-to-End NMS-Free Design:** Building on concepts from [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates Non-Maximum Suppression post-processing. This significantly simplifies deployment code and reduces inference latency variance across all edge devices. 2. **Superior Optimization:** YOLO26 employs the **MuSGD Optimizer**, a hybrid of SGD and Muon (inspired by large language models), which yields highly stable training runs and faster convergence. 3. **Hardware Versatility:** By implementing **DFL Removal** (Distribution Focal Loss), the output heads are simplified, boosting edge device compatibility. In fact, YOLO26 achieves **up to 43% faster CPU inference**, making it vastly superior to YOLOv6 for mobile or IoT edge environments. 4. **Enhanced Accuracy:** Utilizing **ProgLoss + STAL**, YOLO26 sees dramatic improvements in [small object detection](https://www.ultralytics.com/blog/exploring-small-object-detection-with-ultralytics-yolo11), making it the optimal choice for [aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision) and defect inspection. -5. **Unmatched Versatility:** Unlike industrial models that only do bounding boxes, the YOLO26 family supports multi-modal tasks, including [Image Classification](https://docs.ultralytics.com/tasks/classify/), [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +5. **Unmatched Versatility:** Unlike industrial models that only do bounding boxes, the YOLO26 family supports multi-modal tasks, including [Image Classification](https://docs.ultralytics.com/tasks/classify), [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } ### Seamless Ecosystem Experience -The [Ultralytics Platform](https://platform.ultralytics.com) transforms the entire machine learning lifecycle. Training a model is no longer a multi-stage distillation headache. With automatic data augmentation, unified hyperparameter tuning, and one-click exports to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/), [OpenVINO](https://docs.ultralytics.com/integrations/openvino/), and CoreML, you go from dataset to production in hours, not weeks. +The [Ultralytics Platform](https://platform.ultralytics.com) transforms the entire machine learning lifecycle. Training a model is no longer a multi-stage distillation headache. With automatic data augmentation, unified hyperparameter tuning, and one-click exports to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx), [OpenVINO](https://docs.ultralytics.com/integrations/openvino), and CoreML, you go from dataset to production in hours, not weeks. -Additionally, Ultralytics models are known for their [memory efficiency](https://docs.ultralytics.com/guides/yolo-common-issues/), sidestepping the massive VRAM bottlenecks that plague transformer architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +Additionally, Ultralytics models are known for their [memory efficiency](https://docs.ultralytics.com/guides/yolo-common-issues), sidestepping the massive VRAM bottlenecks that plague transformer architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). ### Quick Start Code Example @@ -149,6 +149,6 @@ model.export(format="engine", dynamic=True) Both DAMO-YOLO and YOLOv6-3.0 are impressive engineering feats that push the boundaries of industrial object detection. However, they are highly specialized tools that often require intricate setups and rigid hardware constraints. -For developers and researchers who demand a perfect **performance balance**, multi-task capabilities, and an actively [well-maintained ecosystem](https://www.ultralytics.com/about), Ultralytics **YOLO26** stands unmatched. By blending LLM-inspired optimizers with a clean, NMS-free architecture, YOLO26 simplifies [AI deployment](https://docs.ultralytics.com/guides/model-deployment-options/) while delivering state-of-the-art accuracy across edge and cloud environments. +For developers and researchers who demand a perfect **performance balance**, multi-task capabilities, and an actively [well-maintained ecosystem](https://www.ultralytics.com/about), Ultralytics **YOLO26** stands unmatched. By blending LLM-inspired optimizers with a clean, NMS-free architecture, YOLO26 simplifies [AI deployment](https://docs.ultralytics.com/guides/model-deployment-options) while delivering state-of-the-art accuracy across edge and cloud environments. -If you're evaluating models for a new computer vision project, we highly recommend exploring the capabilities of the [Ultralytics YOLO](https://www.ultralytics.com/yolo) ecosystem. You may also find it useful to compare these with other architectures like [EfficientDet](https://docs.ultralytics.com/compare/efficientdet-vs-yolov6/) or previous milestones like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) to fully grasp the evolution of real-time vision AI. +If you're evaluating models for a new computer vision project, we highly recommend exploring the capabilities of the [Ultralytics YOLO](https://www.ultralytics.com/yolo) ecosystem. You may also find it useful to compare these with other architectures like [EfficientDet](https://docs.ultralytics.com/compare/efficientdet-vs-yolov6) or previous milestones like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) to fully grasp the evolution of real-time vision AI. diff --git a/docs/en/compare/damo-yolo-vs-yolov7.md b/docs/en/compare/damo-yolo-vs-yolov7.md index 7eeffc73667..73224770ce4 100644 --- a/docs/en/compare/damo-yolo-vs-yolov7.md +++ b/docs/en/compare/damo-yolo-vs-yolov7.md @@ -6,7 +6,7 @@ keywords: DAMO-YOLO, YOLOv7, object detection, model comparison, computer vision # DAMO-YOLO vs YOLOv7: Evaluating Real-Time Object Detectors -The rapid evolution of computer vision has produced highly efficient [object detection](https://docs.ultralytics.com/tasks/detect/) models designed to balance precision and computational cost. Two notable models introduced in 2022 are **DAMO-YOLO** and **YOLOv7**. While both aim to push the boundaries of real-time vision tasks, they achieve their results through vastly different architectural paradigms and training methodologies. +The rapid evolution of computer vision has produced highly efficient [object detection](https://docs.ultralytics.com/tasks/detect) models designed to balance precision and computational cost. Two notable models introduced in 2022 are **DAMO-YOLO** and **YOLOv7**. While both aim to push the boundaries of real-time vision tasks, they achieve their results through vastly different architectural paradigms and training methodologies. This comprehensive technical comparison explores the distinct approaches of both models, examining their architectures, deployment potential, and performance metrics to help machine learning engineers choose the right tool for their specific [computer vision applications](https://www.ultralytics.com/blog/60-impactful-computer-vision-applications). @@ -39,9 +39,9 @@ Released as the state-of-the-art in mid-2022, YOLOv7 pushed [real-time inference - **Organization:** [Institute of Information Science, Academia Sinica, Taiwan](https://www.iis.sinica.edu.tw/zh/index.html) - **Date:** July 6, 2022 - **Arxiv:** [2207.02696](https://arxiv.org/abs/2207.02696) -- **Docs:** [YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7/) +- **Docs:** [YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7) -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } !!! tip "Supported Ecosystem" @@ -86,9 +86,9 @@ As seen in the metrics, while DAMO-YOLO provides extremely lightweight variants While theoretical architecture is important, the practicality of a model is dictated by its ecosystem. Models supported by Ultralytics, such as YOLOv7, benefit from a **well-maintained ecosystem** and unparalleled **ease of use**. -- **Performance Balance:** Ultralytics models consistently strike an optimal trade-off between inference speed and detection accuracy, making them ideal for both edge devices and cloud-based [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/). +- **Performance Balance:** Ultralytics models consistently strike an optimal trade-off between inference speed and detection accuracy, making them ideal for both edge devices and cloud-based [model deployment](https://docs.ultralytics.com/guides/model-deployment-options). - **Memory Requirements:** Unlike heavier Transformer-based models, Ultralytics YOLO models maintain low [CUDA](https://developer.nvidia.com/cuda) memory requirements during training. This permits larger [batch sizes](https://www.ultralytics.com/glossary/batch-size), streamlining the training process even on consumer-grade hardware. -- **Versatility:** The Ultralytics framework extends beyond object detection to tasks like [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/) and [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), giving developers a complete computer vision toolkit. +- **Versatility:** The Ultralytics framework extends beyond object detection to tasks like [Instance Segmentation](https://docs.ultralytics.com/tasks/segment) and [Pose Estimation](https://docs.ultralytics.com/tasks/pose), giving developers a complete computer vision toolkit. !!! note "Training Efficiency" @@ -118,11 +118,11 @@ While YOLOv7 and DAMO-YOLO represented significant breakthroughs in 2022, the fi YOLO26 brings a generational leap in performance and usability, incorporating state-of-the-art innovations: -- **End-to-End NMS-Free Design:** YOLO26 is natively end-to-end. By eliminating Non-Maximum Suppression (NMS) post-processing, it delivers faster, simpler deployment logic—a paradigm shift initially pioneered by [YOLOv10](https://docs.ultralytics.com/models/yolov10/). +- **End-to-End NMS-Free Design:** YOLO26 is natively end-to-end. By eliminating Non-Maximum Suppression (NMS) post-processing, it delivers faster, simpler deployment logic—a paradigm shift initially pioneered by [YOLOv10](https://docs.ultralytics.com/models/yolov10). - **MuSGD Optimizer:** Inspired by large language model innovations like Moonshot AI's Kimi K2, YOLO26 utilizes a hybrid of SGD and Muon. This optimizer ensures highly stable training dynamics and dramatically faster convergence rates. - **Up to 43% Faster CPU Inference:** With the targeted removal of Distribution Focal Loss (DFL) and profound structural enhancements, YOLO26 is heavily optimized for low-power edge computing, outperforming previous generations on non-GPU hardware. - **ProgLoss + STAL:** Incorporates advanced new loss functions that explicitly target and improve small-object recognition, an essential capability for applications in aerial imagery, robotics, and [security monitoring](https://www.ultralytics.com/blog/real-time-security-monitoring-with-ai-and-ultralytics-yolo11). -- **Task-Specific Improvements:** Beyond standard detection, YOLO26 features tailored enhancements for diverse tasks, including multi-scale prototyping for segmentation, RLE for pose estimation, and specific angle losses for [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +- **Task-Specific Improvements:** Beyond standard detection, YOLO26 features tailored enhancements for diverse tasks, including multi-scale prototyping for segmentation, RLE for pose estimation, and specific angle losses for [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -143,7 +143,7 @@ Choosing the right architecture depends entirely on your target deployment envir **When to choose YOLO26 (Recommended):** - You are building a new [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) application from scratch and need the absolute state-of-the-art in both precision and CPU/edge inference speed. -- You require rapid, seamless deployment (such as exporting to [CoreML](https://docs.ultralytics.com/integrations/coreml/) or [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/)) without dealing with NMS operator constraints. +- You require rapid, seamless deployment (such as exporting to [CoreML](https://docs.ultralytics.com/integrations/coreml) or [TensorRT](https://docs.ultralytics.com/integrations/tensorrt)) without dealing with NMS operator constraints. - You want to utilize the full capabilities of the [Ultralytics Platform](https://platform.ultralytics.com) for cloud training, dataset management, and automated deployment. By leveraging the robust ecosystem of Ultralytics models, developers can drastically cut down on engineering time while securing top-tier predictive performance for their real-world applications. diff --git a/docs/en/compare/damo-yolo-vs-yolov8.md b/docs/en/compare/damo-yolo-vs-yolov8.md index 25d857c4d4d..7c9d0887cf7 100644 --- a/docs/en/compare/damo-yolo-vs-yolov8.md +++ b/docs/en/compare/damo-yolo-vs-yolov8.md @@ -35,7 +35,7 @@ Understanding the origins of these deep learning models provides valuable contex **Organization:** [Ultralytics](https://www.ultralytics.com/) **Date:** 2023-01-10 **GitHub:** [ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -**Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8/) +**Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8) [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -113,13 +113,13 @@ DAMO-YOLO is an excellent choice for academic environments studying Neural Archi For the vast majority of commercial projects, Ultralytics models provide superior performance balance. -- **Smart Retail:** Using YOLOv8's multi-task capabilities to handle both bounding box detection for inventory and [pose estimation](https://docs.ultralytics.com/tasks/pose/) for analyzing customer behavior. -- **Agriculture:** Employing [instance segmentation](https://docs.ultralytics.com/tasks/segment/) to detect exact plant boundaries and weeds in real-time tractor feeds. -- **Aerial Imagery:** Leveraging [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) to accurately track rotated vehicles and ships from drones or satellites. +- **Smart Retail:** Using YOLOv8's multi-task capabilities to handle both bounding box detection for inventory and [pose estimation](https://docs.ultralytics.com/tasks/pose) for analyzing customer behavior. +- **Agriculture:** Employing [instance segmentation](https://docs.ultralytics.com/tasks/segment) to detect exact plant boundaries and weeds in real-time tractor feeds. +- **Aerial Imagery:** Leveraging [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb) to accurately track rotated vehicles and ships from drones or satellites. !!! note "Other Notable Models" - If you are exploring the broader landscape, you might also be interested in comparing [YOLOv10](https://docs.ultralytics.com/models/yolov10/) or [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) which bring further advancements to anchor-free detection. + If you are exploring the broader landscape, you might also be interested in comparing [YOLOv10](https://docs.ultralytics.com/models/yolov10) or [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) which bring further advancements to anchor-free detection. ## Future-Proofing: Enter YOLO26 diff --git a/docs/en/compare/damo-yolo-vs-yolov9.md b/docs/en/compare/damo-yolo-vs-yolov9.md index a2f260426ab..80af8e20df8 100644 --- a/docs/en/compare/damo-yolo-vs-yolov9.md +++ b/docs/en/compare/damo-yolo-vs-yolov9.md @@ -53,9 +53,9 @@ Introduced as a solution to information loss in deep convolutional networks, YOL - **Release Date:** February 21, 2024 - **Arxiv Paper:** [YOLOv9 Research Paper](https://arxiv.org/abs/2402.13616) - **Official GitHub:** [WongKinYiu/yolov9 Repository](https://github.com/WongKinYiu/yolov9) -- **Documentation:** [YOLOv9 Ultralytics Docs](https://docs.ultralytics.com/models/yolov9/) +- **Documentation:** [YOLOv9 Ultralytics Docs](https://docs.ultralytics.com/models/yolov9) -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ## Architectural Innovations @@ -127,15 +127,15 @@ YOLOv9 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Advancing to YOLO26 -For users comparing legacy architectures, transitioning to the modern Ultralytics ecosystem—specifically the [latest YOLO26 models](https://docs.ultralytics.com/models/yolo26/)—provides an unparalleled advantage. +For users comparing legacy architectures, transitioning to the modern Ultralytics ecosystem—specifically the [latest YOLO26 models](https://docs.ultralytics.com/models/yolo26)—provides an unparalleled advantage. YOLO26 fundamentally alters the deployment landscape through its **End-to-End NMS-Free Design**. By entirely eliminating Non-Maximum Suppression (NMS) post-processing, it delivers faster, dramatically simpler deployment architectures. Coupled with the removal of Distribution Focal Loss (DFL), YOLO26 offers superior compatibility for edge and low-power devices. @@ -158,4 +158,4 @@ results = model.train(data="coco8.yaml", epochs=100, imgsz=640) model.export(format="onnx") ``` -Whether you require advanced [instance segmentation](https://docs.ultralytics.com/tasks/segment/), highly accurate [pose estimation](https://docs.ultralytics.com/tasks/pose/), or standard bounding box detection, the versatility of the Ultralytics framework ensures that your team spends less time configuring deep learning environments and more time deploying robust AI solutions. With specialized task improvements like **ProgLoss + STAL** for enhanced small-object recognition, YOLO26 stands as the premier choice for the next generation of vision applications. +Whether you require advanced [instance segmentation](https://docs.ultralytics.com/tasks/segment), highly accurate [pose estimation](https://docs.ultralytics.com/tasks/pose), or standard bounding box detection, the versatility of the Ultralytics framework ensures that your team spends less time configuring deep learning environments and more time deploying robust AI solutions. With specialized task improvements like **ProgLoss + STAL** for enhanced small-object recognition, YOLO26 stands as the premier choice for the next generation of vision applications. diff --git a/docs/en/compare/damo-yolo-vs-yolox.md b/docs/en/compare/damo-yolo-vs-yolox.md index ef66cc1c95a..2621f7773f7 100644 --- a/docs/en/compare/damo-yolo-vs-yolox.md +++ b/docs/en/compare/damo-yolo-vs-yolox.md @@ -108,7 +108,7 @@ Released in January 2026, YOLO26 is the ultimate recommended model for all [comp - **Up to 43% Faster CPU Inference:** By strategically removing Distribution Focal Loss (DFL) and optimizing the layers, YOLO26 delivers unparalleled speeds on CPUs and edge hardware. - **MuSGD Optimizer:** Inspired by large language model (LLM) training techniques, YOLO26 introduces the MuSGD optimizer (a hybrid of SGD and Muon), resulting in highly stable training runs and much faster convergence compared to the legacy setups in YOLOX. - **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, making YOLO26 vastly superior for drone footage and robotics. -- **Versatility:** Unlike DAMO-YOLO, which is strictly for object detection, YOLO26 seamlessly handles [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) natively within the same well-maintained ecosystem. +- **Versatility:** Unlike DAMO-YOLO, which is strictly for object detection, YOLO26 seamlessly handles [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb) natively within the same well-maintained ecosystem. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -140,6 +140,6 @@ model.export(format="onnx") Choosing between DAMO-YOLO and YOLOX depends on specific constraints: DAMO-YOLO offers exceptional speed-to-accuracy ratios on specific GPUs via NAS, while YOLOX provides a clean, anchor-free design ideal for lightweight edge scenarios. -However, for teams seeking a modern, future-proof solution with an active community, the [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) architecture is the definitive choice. Its NMS-free design, rapid CPU inference, and unified API for detection, segmentation, and pose tasks make it unparalleled for transitioning smoothly from research to robust real-world production. +However, for teams seeking a modern, future-proof solution with an active community, the [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) architecture is the definitive choice. Its NMS-free design, rapid CPU inference, and unified API for detection, segmentation, and pose tasks make it unparalleled for transitioning smoothly from research to robust real-world production. -For developers interested in exploring other modern architectures, we also recommend checking out [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) available in the comprehensive Ultralytics documentation. +For developers interested in exploring other modern architectures, we also recommend checking out [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) available in the comprehensive Ultralytics documentation. diff --git a/docs/en/compare/efficientdet-vs-damo-yolo.md b/docs/en/compare/efficientdet-vs-damo-yolo.md index 94febbde4d2..84af1721336 100644 --- a/docs/en/compare/efficientdet-vs-damo-yolo.md +++ b/docs/en/compare/efficientdet-vs-damo-yolo.md @@ -8,7 +8,7 @@ keywords: EfficientDet, DAMO-YOLO, object detection, model comparison, Efficient When building scalable [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) pipelines, selecting the right model architecture is a critical decision that influences both deployment feasibility and detection accuracy. This guide provides an in-depth, technical comparison between two well-known architectures in the visual recognition landscape: EfficientDet and DAMO-YOLO. -While both models brought significant innovations to the field of [object detection](https://www.ultralytics.com/glossary/object-detection), the rapid advancement of vision AI has paved the way for more integrated ecosystems. Throughout this analysis, we will explore the core mechanics of these legacy networks while illustrating why modern solutions like the [Ultralytics Platform](https://docs.ultralytics.com/platform/) and [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) have become the industry standard for production environments. +While both models brought significant innovations to the field of [object detection](https://www.ultralytics.com/glossary/object-detection), the rapid advancement of vision AI has paved the way for more integrated ecosystems. Throughout this analysis, we will explore the core mechanics of these legacy networks while illustrating why modern solutions like the [Ultralytics Platform](https://docs.ultralytics.com/platform) and [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) have become the industry standard for production environments. @@ -32,7 +32,7 @@ EfficientDet's primary contribution is the Bi-directional Feature Pyramid Networ ### Strengths and Weaknesses -The key strength of EfficientDet lies in its parameter efficiency. For tasks where [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map) needs to be maximized on heavily constrained cloud environments, its compound scaling method is highly predictable. However, EfficientDet is notoriously complex to train from scratch and often demands substantial [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/). Furthermore, its heavy reliance on specific TensorFlow operations makes transitioning to edge deployments via ONNX or TensorRT more cumbersome compared to the streamlined [export capabilities](https://docs.ultralytics.com/modes/export/) found in modern YOLO models. +The key strength of EfficientDet lies in its parameter efficiency. For tasks where [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map) needs to be maximized on heavily constrained cloud environments, its compound scaling method is highly predictable. However, EfficientDet is notoriously complex to train from scratch and often demands substantial [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning). Furthermore, its heavy reliance on specific TensorFlow operations makes transitioning to edge deployments via ONNX or TensorRT more cumbersome compared to the streamlined [export capabilities](https://docs.ultralytics.com/modes/export) found in modern YOLO models. [Learn more about EfficientDet](https://github.com/google/automl/tree/master/efficientdet#readme){ .md-button } @@ -53,13 +53,13 @@ DAMO-YOLO introduces several novel technologies. It utilizes a NAS-generated bac ### Strengths and Weaknesses -DAMO-YOLO shines in its GPU inference speeds, specifically engineered for deployment on NVIDIA architectures using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). By stripping away heavy head structures, the model delivers low-latency predictions. Conversely, the automated architecture search can make the model structure opaque and difficult to manually debug or fine-tune for custom edge devices. Unlike the highly versatile [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), DAMO-YOLO is primarily focused on standard bounding box detection, lacking native support for advanced tasks like [pose estimation](https://docs.ultralytics.com/tasks/pose/) or [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection out of the box. +DAMO-YOLO shines in its GPU inference speeds, specifically engineered for deployment on NVIDIA architectures using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). By stripping away heavy head structures, the model delivers low-latency predictions. Conversely, the automated architecture search can make the model structure opaque and difficult to manually debug or fine-tune for custom edge devices. Unlike the highly versatile [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), DAMO-YOLO is primarily focused on standard bounding box detection, lacking native support for advanced tasks like [pose estimation](https://docs.ultralytics.com/tasks/pose) or [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb) detection out of the box. [Learn more about DAMO-YOLO](https://github.com/tinyvision/DAMO-YOLO/blob/master/README.md){ .md-button } ## Performance Comparison -Understanding the empirical trade-offs is essential for selecting a model. The table below compares the EfficientDet family against the DAMO-YOLO series across crucial [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/). +Understanding the empirical trade-offs is essential for selecting a model. The table below compares the EfficientDet family against the DAMO-YOLO series across crucial [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | --------------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -103,11 +103,11 @@ DAMO-YOLO is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Advancing Beyond Legacy Models @@ -115,7 +115,7 @@ While EfficientDet and DAMO-YOLO provide valuable academic insights, modern deve ### Unmatched Ease of Use and Ecosystem -Deploying models from separate, heavily customized research repositories often leads to integration nightmares. Ultralytics provides a unified, deeply [well-maintained ecosystem](https://docs.ultralytics.com/help/contributing/) with extensive documentation and a pythonic API. Whether you are using [Google Colab](https://docs.ultralytics.com/integrations/google-colab/) for training or exporting to [CoreML](https://docs.ultralytics.com/integrations/coreml/) for mobile inference, the pipeline requires only a few lines of code. +Deploying models from separate, heavily customized research repositories often leads to integration nightmares. Ultralytics provides a unified, deeply [well-maintained ecosystem](https://docs.ultralytics.com/help/contributing) with extensive documentation and a pythonic API. Whether you are using [Google Colab](https://docs.ultralytics.com/integrations/google-colab) for training or exporting to [CoreML](https://docs.ultralytics.com/integrations/coreml) for mobile inference, the pipeline requires only a few lines of code. ```python from ultralytics import YOLO @@ -134,18 +134,18 @@ model.export(format="onnx") For developers evaluating EfficientDet or DAMO-YOLO, [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) represents the ultimate evolutionary step. Released in early 2026, it introduces paradigm-shifting capabilities: -- **End-to-End NMS-Free Design:** First pioneered by [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates the need for Non-Maximum Suppression (NMS) post-processing. This translates to vastly simpler deployment architectures and consistent latency across diverse hardware. +- **End-to-End NMS-Free Design:** First pioneered by [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates the need for Non-Maximum Suppression (NMS) post-processing. This translates to vastly simpler deployment architectures and consistent latency across diverse hardware. - **Up to 43% Faster CPU Inference:** For edge deployments lacking heavy GPUs—scenarios where DAMO-YOLO struggles—YOLO26 is heavily optimized, delivering massive speedups on standard CPUs. - **MuSGD Optimizer:** Bridging the gap between LLM innovations and computer vision, YOLO26 incorporates the MuSGD optimizer (inspired by Moonshot AI), ensuring incredibly stable training and rapid convergence compared to the brittle training loops of EfficientDet. -- **DFL Removal:** The removal of Distribution Focal Loss simplifies the export process, guaranteeing superior compatibility with low-power microcontrollers and [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) devices. +- **DFL Removal:** The removal of Distribution Focal Loss simplifies the export process, guaranteeing superior compatibility with low-power microcontrollers and [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) devices. - **ProgLoss + STAL:** These advanced loss functions yield dramatic improvements in small-object recognition, an area where older architectures traditionally fail. ### Memory Efficiency and Task Versatility Unlike [transformer](https://www.ultralytics.com/glossary/transformer) models or heavily fused NAS networks, Ultralytics models are characterized by their stringent memory efficiency. They consume remarkably lower CUDA memory during training, enabling rapid iteration on consumer-grade hardware. -Furthermore, while EfficientDet and DAMO-YOLO are rigidly constrained to bounding boxes, Ultralytics natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [image classification](https://docs.ultralytics.com/tasks/classify/) within the exact same intuitive framework. For users maintaining older projects, [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) remains a rock-solid, widely deployed alternative worth exploring. +Furthermore, while EfficientDet and DAMO-YOLO are rigidly constrained to bounding boxes, Ultralytics natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [image classification](https://docs.ultralytics.com/tasks/classify) within the exact same intuitive framework. For users maintaining older projects, [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) remains a rock-solid, widely deployed alternative worth exploring. ## Conclusion -Choosing the right vision architecture involves weighing raw theoretical performance against deployment reality. EfficientDet offers a mathematically elegant scaling approach, and DAMO-YOLO delivers compelling raw GPU speeds. However, for teams prioritizing rapid development, reliable deployments, and cutting-edge features, [Ultralytics models](https://docs.ultralytics.com/models/) stand clearly ahead. By combining innovations like NMS-free inference and MuSGD optimization, [YOLO26](https://docs.ultralytics.com/models/yolo26/) ensures that your computer vision projects are built on the most capable, maintainable, and efficient foundation available today. +Choosing the right vision architecture involves weighing raw theoretical performance against deployment reality. EfficientDet offers a mathematically elegant scaling approach, and DAMO-YOLO delivers compelling raw GPU speeds. However, for teams prioritizing rapid development, reliable deployments, and cutting-edge features, [Ultralytics models](https://docs.ultralytics.com/models) stand clearly ahead. By combining innovations like NMS-free inference and MuSGD optimization, [YOLO26](https://docs.ultralytics.com/models/yolo26) ensures that your computer vision projects are built on the most capable, maintainable, and efficient foundation available today. diff --git a/docs/en/compare/efficientdet-vs-pp-yoloe.md b/docs/en/compare/efficientdet-vs-pp-yoloe.md index 517db8544f5..8eff842c2c6 100644 --- a/docs/en/compare/efficientdet-vs-pp-yoloe.md +++ b/docs/en/compare/efficientdet-vs-pp-yoloe.md @@ -124,10 +124,10 @@ predictions = model("https://ultralytics.com/images/bus.jpg") model.export(format="onnx") ``` -Whether you require standard detection, or specialized tasks like instance segmentation and [pose estimation](https://docs.ultralytics.com/tasks/pose/), YOLO26 supports these natively with multi-scale prototypes and Residual Log-Likelihood Estimation (RLE), all within the exact same user-friendly framework. +Whether you require standard detection, or specialized tasks like instance segmentation and [pose estimation](https://docs.ultralytics.com/tasks/pose), YOLO26 supports these natively with multi-scale prototypes and Residual Log-Likelihood Estimation (RLE), all within the exact same user-friendly framework. ## Exploring Other Notable Models -If you are evaluating architectures for specific enterprise requirements, it is also worth considering the previous generation [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), which remains a robust, production-tested workhorse. For applications where transformer-based architectures are desired, [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) offers an interesting alternative, though it typically demands higher CUDA memory overhead during training compared to the highly efficient YOLO variants. +If you are evaluating architectures for specific enterprise requirements, it is also worth considering the previous generation [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), which remains a robust, production-tested workhorse. For applications where transformer-based architectures are desired, [RT-DETR](https://docs.ultralytics.com/models/rtdetr) offers an interesting alternative, though it typically demands higher CUDA memory overhead during training compared to the highly efficient YOLO variants. In conclusion, while EfficientDet offers principled scaling and PP-YOLOE+ provides excellent GPU throughput within its specific framework, **Ultralytics YOLO26** delivers the most balanced, versatile, and developer-friendly solution available today. Its natively end-to-end architecture and extensive integration capabilities make it the recommended foundation for next-generation vision AI. diff --git a/docs/en/compare/efficientdet-vs-rtdetr.md b/docs/en/compare/efficientdet-vs-rtdetr.md index ad59fe5f777..d2c03e76036 100644 --- a/docs/en/compare/efficientdet-vs-rtdetr.md +++ b/docs/en/compare/efficientdet-vs-rtdetr.md @@ -17,7 +17,7 @@ By understanding the trade-offs between legacy efficiency and modern transformer ## Understanding EfficientDet -EfficientDet revolutionized [object detection](https://docs.ultralytics.com/tasks/detect/) by introducing a principled approach to model scaling. +EfficientDet revolutionized [object detection](https://docs.ultralytics.com/tasks/detect) by introducing a principled approach to model scaling. - **Authors:** Mingxing Tan, Ruoming Pang, and Quoc V. Le - **Organization:** [Google](https://ai.google/) @@ -40,7 +40,7 @@ However, EfficientDet relies on standard non-maximum suppression (NMS) during po !!! info "Legacy Support" - While EfficientDet paved the way for scalable networks, deploying these models on modern NPUs often requires extensive manual optimization. For streamlined deployments, newer [Ultralytics models](https://docs.ultralytics.com/models/) offer 1-click export functionality. + While EfficientDet paved the way for scalable networks, deploying these models on modern NPUs often requires extensive manual optimization. For streamlined deployments, newer [Ultralytics models](https://docs.ultralytics.com/models) offer 1-click export functionality. ## Exploring RTDETRv2 @@ -55,11 +55,11 @@ RTDETRv2 represents the evolution of transformer-based architectures, shifting t ### Advancements in Transformers -[RTDETRv2](https://docs.ultralytics.com/models/rtdetr/) builds upon the Real-Time Detection Transformer (RT-DETR) baseline. It leverages global attention mechanisms, enabling the model to understand complex scene contexts without the localized constraints of standard convolutions. The most significant architectural advantage is its natively NMS-free design. By predicting objects directly from the input image, it simplifies the inference pipeline, avoiding the heuristic tuning required by NMS post-processing. +[RTDETRv2](https://docs.ultralytics.com/models/rtdetr) builds upon the Real-Time Detection Transformer (RT-DETR) baseline. It leverages global attention mechanisms, enabling the model to understand complex scene contexts without the localized constraints of standard convolutions. The most significant architectural advantage is its natively NMS-free design. By predicting objects directly from the input image, it simplifies the inference pipeline, avoiding the heuristic tuning required by NMS post-processing. ### Strengths and Weaknesses -RTDETRv2 excels in high-density environments where overlapping objects confuse traditional CNNs. It is highly accurate on complex benchmark [datasets like COCO](https://docs.ultralytics.com/datasets/detect/coco/). +RTDETRv2 excels in high-density environments where overlapping objects confuse traditional CNNs. It is highly accurate on complex benchmark [datasets like COCO](https://docs.ultralytics.com/datasets/detect/coco). Despite its accuracy, transformer models naturally demand substantial memory. The training efficiency is notably lower; it requires significantly more epochs and higher [CUDA](https://developer.nvidia.com/cuda/toolkit) memory footprints to converge compared to CNNs. This makes RTDETRv2 less ideal for developers operating with constrained cloud budgets or those needing rapid rapid prototyping. @@ -71,7 +71,7 @@ Despite its accuracy, transformer models naturally demand substantial memory. Th ## Performance Benchmark Comparison -Understanding the raw [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/) is vital for model selection. The following table showcases the comparison between EfficientDet and RTDETRv2 across various sizes. +Understanding the raw [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics) is vital for model selection. The following table showcases the comparison between EfficientDet and RTDETRv2 across various sizes. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | --------------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -111,11 +111,11 @@ RT-DETR is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Introducing YOLO26 @@ -128,9 +128,9 @@ YOLO26 stands out by combining the streamlined ecosystem [Ultralytics](https://w - **End-to-End NMS-Free Design:** Taking inspiration from transformers like RTDETRv2, YOLO26 is natively end-to-end. It eliminates NMS post-processing, guaranteeing faster, simpler deployment pipelines without the massive parameter bloat of pure transformers. - **MuSGD Optimizer:** Inspired by large language model training innovations (like Moonshot AI's Kimi K2), YOLO26 utilizes a hybrid of SGD and Muon. This brings unprecedented training stability and significantly faster convergence rates compared to the prolonged schedules required by RTDETRv2. - **Optimized for Edge:** With up to **43% faster CPU inference**, YOLO26 is built for [edge AI](https://www.ultralytics.com/glossary/edge-ai). It easily outperforms heavy transformer models on constrained hardware like mobile phones and smart cameras. -- **DFL Removal:** The removal of Distribution Focal Loss simplifies the model graph, facilitating seamless [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) and [ONNX](https://docs.ultralytics.com/integrations/onnx/) exports. +- **DFL Removal:** The removal of Distribution Focal Loss simplifies the model graph, facilitating seamless [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) and [ONNX](https://docs.ultralytics.com/integrations/onnx) exports. - **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, solving a common bottleneck in aerial imagery and robotics. -- **Versatility:** Unlike RTDETRv2, which primarily focuses on detection, YOLO26 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) with task-specific improvements like RLE for pose and specialized angle loss for OBB. +- **Versatility:** Unlike RTDETRv2, which primarily focuses on detection, YOLO26 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb) with task-specific improvements like RLE for pose and specialized angle loss for OBB. !!! tip "Integrated Ecosystem" @@ -138,7 +138,7 @@ YOLO26 stands out by combining the streamlined ecosystem [Ultralytics](https://w ### Code Simplicity with Ultralytics -The well-maintained [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) makes model training and inference trivial. Developers can easily benchmark models or launch training scripts with minimal boilerplate code. +The well-maintained [Ultralytics Python API](https://docs.ultralytics.com/usage/python) makes model training and inference trivial. Developers can easily benchmark models or launch training scripts with minimal boilerplate code. ```python from ultralytics import YOLO @@ -153,4 +153,4 @@ results = model.train(data="coco8.yaml", epochs=50, imgsz=640) predictions = model.predict("image.jpg") ``` -For those managing legacy infrastructure, the highly acclaimed [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) remains a stable and powerful choice, showcasing the long-term reliability of the Ultralytics ecosystem. Whether you are running complex [real-time tracking](https://docs.ultralytics.com/modes/track/) algorithms or simple defect detection, upgrading to YOLO26 ensures your system is future-proof, highly accurate, and memory-efficient. +For those managing legacy infrastructure, the highly acclaimed [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) remains a stable and powerful choice, showcasing the long-term reliability of the Ultralytics ecosystem. Whether you are running complex [real-time tracking](https://docs.ultralytics.com/modes/track) algorithms or simple defect detection, upgrading to YOLO26 ensures your system is future-proof, highly accurate, and memory-efficient. diff --git a/docs/en/compare/efficientdet-vs-yolo11.md b/docs/en/compare/efficientdet-vs-yolo11.md index c0ef673cc5b..08a24793973 100644 --- a/docs/en/compare/efficientdet-vs-yolo11.md +++ b/docs/en/compare/efficientdet-vs-yolo11.md @@ -17,7 +17,7 @@ Whether you are targeting millisecond latency on [edge AI](https://www.ultralyti ## Model Profiles and Technical Details -Understanding the lineage and underlying design philosophy of each architecture helps contextualize their performance in real-world [object detection](https://docs.ultralytics.com/tasks/detect/) tasks. +Understanding the lineage and underlying design philosophy of each architecture helps contextualize their performance in real-world [object detection](https://docs.ultralytics.com/tasks/detect) tasks. ### EfficientDet @@ -40,7 +40,7 @@ YOLO11 represents a significant evolution in the Ultralytics ecosystem, pushing - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2024-09-27 - **GitHub:** [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11/) +- **Docs:** [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11) [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } @@ -54,7 +54,7 @@ YOLO11, on the other hand, utilizes an optimized C2f module and an advanced anch !!! tip "Multi-Task Versatility" - While EfficientDet is strictly an object detector, YOLO11 boasts extreme versatility. A single YOLO11 architecture natively supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). + While EfficientDet is strictly an object detector, YOLO11 boasts extreme versatility. A single YOLO11 architecture natively supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Image Classification](https://docs.ultralytics.com/tasks/classify), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). ## Performance Benchmarks @@ -89,11 +89,11 @@ The developer experience is often as critical as the model's theoretical capabil EfficientDet relies heavily on the legacy [TensorFlow](https://www.tensorflow.org/) ecosystem and complex AutoML libraries. Setting up a custom training pipeline involves steep learning curves, intricate dependency management, and manual configuration of anchors and [loss functions](https://www.ultralytics.com/glossary/loss-function). -Conversely, Ultralytics offers an unparalleled ease of use. Backed by a well-maintained PyTorch ecosystem, training a YOLO model requires just a few lines of code. The framework automatically manages [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/), advanced data augmentations, and optimal learning rate scheduling out of the box. +Conversely, Ultralytics offers an unparalleled ease of use. Backed by a well-maintained PyTorch ecosystem, training a YOLO model requires just a few lines of code. The framework automatically manages [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning), advanced data augmentations, and optimal learning rate scheduling out of the box. ### Code Example: Getting Started with Ultralytics -This robust, production-ready snippet demonstrates how straightforward training and inference are within the [Python API](https://docs.ultralytics.com/usage/python/). +This robust, production-ready snippet demonstrates how straightforward training and inference are within the [Python API](https://docs.ultralytics.com/usage/python). ```python from ultralytics import YOLO @@ -115,15 +115,15 @@ results[0].show() EfficientDet remains a viable choice for research environments heavily entrenched in TensorFlow pipelines or specific CPU-bound constraints where early architectures like d0 perform adequately. **When to use YOLO11:** -YOLO11 is the definitive choice for modern enterprise deployments. Its exceptional speed makes it perfect for [autonomous vehicles](https://www.ultralytics.com/blog/ai-in-self-driving-cars), real-time sports analytics, and high-throughput manufacturing defect detection. Furthermore, its lower memory usage enables flexible deployment on resource-constrained hardware like the [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/). +YOLO11 is the definitive choice for modern enterprise deployments. Its exceptional speed makes it perfect for [autonomous vehicles](https://www.ultralytics.com/blog/ai-in-self-driving-cars), real-time sports analytics, and high-throughput manufacturing defect detection. Furthermore, its lower memory usage enables flexible deployment on resource-constrained hardware like the [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson). ## Looking Forward: The YOLO26 Upgrade While YOLO11 is exceptionally capable, developers starting new projects should evaluate other Ultralytics architectures like the proven [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) or the newly released [YOLO26](https://platform.ultralytics.com/ultralytics/yolo26). Released in early 2026, YOLO26 takes the foundation of YOLO11 and introduces several groundbreaking innovations: -- **End-to-End NMS-Free Design:** Building on the legacy of [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 completely eliminates Non-Maximum Suppression (NMS) during post-processing, slashing latency and simplifying deployment pipelines. +- **End-to-End NMS-Free Design:** Building on the legacy of [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 completely eliminates Non-Maximum Suppression (NMS) during post-processing, slashing latency and simplifying deployment pipelines. - **MuSGD Optimizer:** A hybrid optimizer blending standard SGD with Muon (inspired by large language model training), drastically improving training stability. - **Up to 43% Faster CPU Inference:** Specific optimizations make YOLO26 incredibly potent on edge devices lacking discrete GPUs. - **ProgLoss + STAL:** Advanced loss functions that remarkably improve small-object detection, critical for aerial imagery and robotics. -Explore the broader landscape of vision architectures, including transformer-based detectors like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/), in our comprehensive [Ultralytics Docs](https://docs.ultralytics.com/). +Explore the broader landscape of vision architectures, including transformer-based detectors like [RT-DETR](https://docs.ultralytics.com/models/rtdetr), in our comprehensive [Ultralytics Docs](https://docs.ultralytics.com/). diff --git a/docs/en/compare/efficientdet-vs-yolo26.md b/docs/en/compare/efficientdet-vs-yolo26.md index ae24ae59655..7d269c9362d 100644 --- a/docs/en/compare/efficientdet-vs-yolo26.md +++ b/docs/en/compare/efficientdet-vs-yolo26.md @@ -40,16 +40,16 @@ The differences in architecture between these two models are stark, reflecting t EfficientDet was built around the BiFPN (Bi-directional Feature Pyramid Network) and utilizes a compound scaling method across resolution, depth, and width. While it achieved excellent theoretical efficiency in 2019, it relies heavily on legacy TensorFlow frameworks and complex AutoML search algorithms that are often cumbersome to adapt for custom datasets. -In contrast, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) represents the absolute cutting edge of real-time computer vision. It introduces several groundbreaking architectural improvements designed specifically for modern deployment pipelines: +In contrast, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) represents the absolute cutting edge of real-time computer vision. It introduces several groundbreaking architectural improvements designed specifically for modern deployment pipelines: - **End-to-End NMS-Free Design:** YOLO26 is natively end-to-end, completely eliminating the need for Non-Maximum Suppression (NMS) post-processing. This breakthrough approach, first pioneered in [YOLOv10](https://platform.ultralytics.com/ultralytics/yolov10), ensures faster, simpler deployment logic and drastically reduces latency variance on edge chips. - **DFL Removal:** By removing the Distribution Focal Loss (DFL), YOLO26 simplifies the output head, leading to superior compatibility with edge computing and low-power devices. - **MuSGD Optimizer:** Inspired by large language model innovations like Moonshot AI's Kimi K2, YOLO26 utilizes the MuSGD optimizer—a hybrid of SGD and Muon. This delivers dramatically more stable training and faster convergence than standard optimizers. -- **ProgLoss + STAL:** The introduction of Progressive Loss combined with Scale-aware Task-aligned Learning (STAL) provides notable improvements in small-object recognition, which is highly critical for [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and robotics. +- **ProgLoss + STAL:** The introduction of Progressive Loss combined with Scale-aware Task-aligned Learning (STAL) provides notable improvements in small-object recognition, which is highly critical for [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and robotics. !!! tip "Pro Tip: NMS-Free Deployment" - Because YOLO26 eliminates NMS, the entire model can be executed as a single, continuous compute graph. This makes exporting to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/) or [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) incredibly straightforward and maximizes NPU/GPU utilization. + Because YOLO26 eliminates NMS, the entire model can be executed as a single, continuous compute graph. This makes exporting to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx) or [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) incredibly straightforward and maximizes NPU/GPU utilization. ## Performance Metrics and Benchmarks @@ -78,17 +78,17 @@ As demonstrated above, YOLO26 offers a vastly superior **Performance Balance**. Choosing an architecture is rarely just about theoretical FLOPs; it is heavily dependent on the engineering workflows. Developers routinely favor Ultralytics due to the unmatched **Ease of Use**. -EfficientDet training often requires complex dependency management, manual hyperparameter tuning, and legacy TensorFlow setups. Conversely, [Ultralytics models](https://docs.ultralytics.com/models/) feature an elegantly simple API. This seamless experience extends directly into the [Ultralytics Platform](https://platform.ultralytics.com/), which handles cloud training, data annotation, and real-time experiment tracking out-of-the-box. +EfficientDet training often requires complex dependency management, manual hyperparameter tuning, and legacy TensorFlow setups. Conversely, [Ultralytics models](https://docs.ultralytics.com/models) feature an elegantly simple API. This seamless experience extends directly into the [Ultralytics Platform](https://platform.ultralytics.com/), which handles cloud training, data annotation, and real-time experiment tracking out-of-the-box. Furthermore, transformer-based detectors and complex AutoML models suffer from exorbitant memory consumption. Ultralytics models are renowned for their highly efficient **Memory Requirements**, meaning you can train robust models on consumer-grade hardware without encountering out-of-memory (OOM) errors. ### Versatility and Task Support -EfficientDet is strictly an [object detection](https://docs.ultralytics.com/tasks/detect/) network. YOLO26 is a unified multi-task learner. It includes task-specific innovations natively built into the architecture: +EfficientDet is strictly an [object detection](https://docs.ultralytics.com/tasks/detect) network. YOLO26 is a unified multi-task learner. It includes task-specific innovations natively built into the architecture: -- Semantic segmentation loss and multi-scale proto for flawless [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/). -- Residual Log-Likelihood Estimation (RLE) to drastically improve [Pose Estimation](https://docs.ultralytics.com/tasks/pose/) accuracy. -- Specialized angle loss routines for solving boundary issues in [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +- Semantic segmentation loss and multi-scale proto for flawless [Instance Segmentation](https://docs.ultralytics.com/tasks/segment). +- Residual Log-Likelihood Estimation (RLE) to drastically improve [Pose Estimation](https://docs.ultralytics.com/tasks/pose) accuracy. +- Specialized angle loss routines for solving boundary issues in [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). !!! note "Legacy Support" @@ -112,7 +112,7 @@ YOLO26 is recommended for: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Implementation Example: Training YOLO26 @@ -139,6 +139,6 @@ print(f"Model seamlessly exported to: {exported_path}") ## Conclusion: Which Model Should You Choose? -When comparing EfficientDet and YOLO26, the trajectory of the industry is clear. EfficientDet remains an important historical stepping stone in compound scaling research. However, for modern applications—whether deployed on cloud clusters or constrained [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) devices—the choice is heavily skewed toward Ultralytics. +When comparing EfficientDet and YOLO26, the trajectory of the industry is clear. EfficientDet remains an important historical stepping stone in compound scaling research. However, for modern applications—whether deployed on cloud clusters or constrained [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) devices—the choice is heavily skewed toward Ultralytics. By eliminating NMS, optimizing for drastically lower VRAM, and wrapping the technology in a world-class developer ecosystem, YOLO26 is definitively the recommended architecture for robust, production-ready computer vision. Whether you are detecting manufacturing defects or mapping agricultural yields, the [Ultralytics Platform](https://platform.ultralytics.com/) ensures you get from dataset to deployment with unrivaled speed and accuracy. diff --git a/docs/en/compare/efficientdet-vs-yolov10.md b/docs/en/compare/efficientdet-vs-yolov10.md index 649079c3c5e..90096b8da21 100644 --- a/docs/en/compare/efficientdet-vs-yolov10.md +++ b/docs/en/compare/efficientdet-vs-yolov10.md @@ -8,7 +8,7 @@ keywords: EfficientDet,YOLOv10,object detection,model comparison,computer vision In the rapidly evolving field of [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv), choosing the right object detection architecture is critical for balancing accuracy, latency, and computational efficiency. This comprehensive technical guide compares two highly influential models: Google's **EfficientDet** and Tsinghua University's **YOLOv10**. While both models represent significant leaps in object detection, they approach architectural design and [model optimization](https://www.ultralytics.com/blog/what-is-model-optimization-a-quick-guide) from vastly different angles. -We will explore their core architectures, review performance benchmarks on [standard datasets like COCO](https://docs.ultralytics.com/datasets/detect/coco/), and discuss how they integrate into modern machine learning pipelines, specifically highlighting the advantages of the comprehensive [Ultralytics ecosystem](https://www.ultralytics.com). +We will explore their core architectures, review performance benchmarks on [standard datasets like COCO](https://docs.ultralytics.com/datasets/detect/coco), and discuss how they integrate into modern machine learning pipelines, specifically highlighting the advantages of the comprehensive [Ultralytics ecosystem](https://www.ultralytics.com). @@ -47,7 +47,7 @@ Released in mid-2024, YOLOv10 fundamentally changed the real-time object detecti YOLOv10 introduces a consistent dual-assignment strategy for NMS-free training. By utilizing both one-to-many and one-to-one label assignments during training, the network learns to produce uniquely matching bounding boxes without relying on NMS to filter out duplicates. This holistic efficiency-accuracy driven model design reduces computational redundancy, making it an excellent candidate for [edge computing](https://www.ultralytics.com/glossary/edge-computing) and low-latency video streaming applications. It seamlessly integrates into the Ultralytics ecosystem, granting developers access to an extremely straightforward Python API. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } !!! info "NMS-Free Impact" @@ -79,10 +79,10 @@ _Note: The YOLOv10n variant requires significantly fewer parameters (2.3M) and a ## Why Choose Ultralytics for Model Deployment? -While both models have historical and structural significance, integrating them into modern pipelines can be a challenge. This is where the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolov10) shines. By providing a unified ecosystem, Ultralytics simplifies the entire lifecycle—from [data annotation](https://docs.ultralytics.com/platform/data/annotation/) to deployment. +While both models have historical and structural significance, integrating them into modern pipelines can be a challenge. This is where the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolov10) shines. By providing a unified ecosystem, Ultralytics simplifies the entire lifecycle—from [data annotation](https://docs.ultralytics.com/platform/data/annotation) to deployment. -1. **Ease of Use:** The Ultralytics Python package offers a single interface for [model training](https://docs.ultralytics.com/modes/train/), [validation](https://docs.ultralytics.com/modes/val/), and export, replacing hundreds of lines of boilerplate code with concise commands. -2. **Ecosystem and Versatility:** While EfficientDet is heavily specialized for detection, Ultralytics YOLO models naturally extend to [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), and Classification. +1. **Ease of Use:** The Ultralytics Python package offers a single interface for [model training](https://docs.ultralytics.com/modes/train), [validation](https://docs.ultralytics.com/modes/val), and export, replacing hundreds of lines of boilerplate code with concise commands. +2. **Ecosystem and Versatility:** While EfficientDet is heavily specialized for detection, Ultralytics YOLO models naturally extend to [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb), and Classification. 3. **Training Efficiency:** Leveraging cutting-edge techniques like auto-batching and distributed training, Ultralytics models train faster and consume drastically less CUDA memory than heavy transformer or older multi-branch TF architectures. ### Code Example: Training YOLOv10 @@ -127,15 +127,15 @@ YOLOv10 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Future is Here: Enter Ultralytics YOLO26 -While YOLOv10 introduced the revolutionary NMS-free design, the technology has evolved. Released in January 2026, [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) represents the definitive state-of-the-art for vision AI. It unifies the best aspects of previous architectures—like the [YOLO11](https://docs.ultralytics.com/models/yolo11/) multi-task capabilities and [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) stability—into a singular, highly optimized powerhouse. +While YOLOv10 introduced the revolutionary NMS-free design, the technology has evolved. Released in January 2026, [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) represents the definitive state-of-the-art for vision AI. It unifies the best aspects of previous architectures—like the [YOLO11](https://docs.ultralytics.com/models/yolo11) multi-task capabilities and [RT-DETR](https://docs.ultralytics.com/models/rtdetr) stability—into a singular, highly optimized powerhouse. !!! tip "The YOLO26 Advantage" @@ -156,4 +156,4 @@ Choosing between these networks ultimately depends on your deployment constraint - **EfficientDet** remains a topic of academic interest regarding compound scaling and is suitable for researchers maintaining existing [TensorFlow](https://www.ultralytics.com/glossary/tensorflow) systems where model weight size (on disk) is more critical than runtime speed. - **YOLOv10** is phenomenal for applications demanding ultra-low latency, such as high-speed [multi-object tracking](https://www.ultralytics.com/glossary/multi-object-tracking-mot) and traffic monitoring, due to its pioneering NMS-free architecture. -- **YOLO26**, however, is the ultimate recommendation for modern [computer vision projects](https://docs.ultralytics.com/guides/steps-of-a-cv-project/), offering the absolute highest [Performance Balance](https://docs.ultralytics.com/guides/yolo-performance-metrics/) of accuracy, minimal memory footprint, and multi-task versatility supported by the robust Ultralytics ecosystem. +- **YOLO26**, however, is the ultimate recommendation for modern [computer vision projects](https://docs.ultralytics.com/guides/steps-of-a-cv-project), offering the absolute highest [Performance Balance](https://docs.ultralytics.com/guides/yolo-performance-metrics) of accuracy, minimal memory footprint, and multi-task versatility supported by the robust Ultralytics ecosystem. diff --git a/docs/en/compare/efficientdet-vs-yolov5.md b/docs/en/compare/efficientdet-vs-yolov5.md index 311077a9aad..7a046fb3b28 100644 --- a/docs/en/compare/efficientdet-vs-yolov5.md +++ b/docs/en/compare/efficientdet-vs-yolov5.md @@ -29,7 +29,7 @@ Introduced by Google Research, EfficientDet was designed to systematically scale ### Architectural Innovations -EfficientDet leverages the EfficientNet classification model as its backbone, utilizing a compound scaling method that uniformly scales network width, depth, and resolution. Its most notable contribution to [object detection](https://docs.ultralytics.com/tasks/detect/) is the introduction of the Bi-directional Feature Pyramid Network (BiFPN). Unlike standard Feature Pyramid Networks that simply aggregate features top-down, BiFPN allows for complex, bidirectional cross-scale connections and introduces learnable weights to determine the importance of different input features. +EfficientDet leverages the EfficientNet classification model as its backbone, utilizing a compound scaling method that uniformly scales network width, depth, and resolution. Its most notable contribution to [object detection](https://docs.ultralytics.com/tasks/detect) is the introduction of the Bi-directional Feature Pyramid Network (BiFPN). Unlike standard Feature Pyramid Networks that simply aggregate features top-down, BiFPN allows for complex, bidirectional cross-scale connections and introduces learnable weights to determine the importance of different input features. While highly accurate, EfficientDet relies heavily on the [TensorFlow](https://www.tensorflow.org/) ecosystem and specific AutoML libraries. This dependency can sometimes make it cumbersome to integrate into custom, lightweight deployment pipelines or environments that favor dynamic computational graphs. @@ -45,19 +45,19 @@ Released shortly after EfficientDet, [Ultralytics YOLOv5](https://platform.ultra - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** June 26, 2020 - **GitHub:** [ultralytics/yolov5](https://github.com/ultralytics/yolov5) -- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5/) +- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5) ### Architectural Innovations YOLOv5 introduced significant upgrades over its predecessors, utilizing a CSPDarknet (Cross-Stage Partial) backbone that significantly enhances gradient flow while reducing the overall parameter count. Furthermore, YOLOv5 incorporates Auto-Learning Anchor Boxes, which automatically calculate the optimal bounding box priors based on your specific custom training data, eliminating the need for manual hyperparameter tuning. -YOLOv5 also heavily utilizes [Mosaic Data Augmentation](https://docs.ultralytics.com/reference/data/augment/), blending four disparate images into a single training tile. This greatly improves the model's ability to detect small objects and generalizes contextual understanding, making it highly robust in varied environments. +YOLOv5 also heavily utilizes [Mosaic Data Augmentation](https://docs.ultralytics.com/reference/data/augment), blending four disparate images into a single training tile. This greatly improves the model's ability to detect small objects and generalizes contextual understanding, making it highly robust in varied environments. [Learn more about YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5){ .md-button } ## Performance and Benchmarks -Evaluating models on standard benchmarks like the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/) is crucial for understanding the trade-offs between precision and speed. The table below illustrates how different sizes of EfficientDet and YOLOv5 perform under standardized conditions. +Evaluating models on standard benchmarks like the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco) is crucial for understanding the trade-offs between precision and speed. The table below illustrates how different sizes of EfficientDet and YOLOv5 perform under standardized conditions. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | --------------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -84,13 +84,13 @@ Additionally, YOLOv5 models demonstrate much lower CUDA memory requirements duri !!! tip "Maximizing Hardware Efficiency" - To extract the maximum frames-per-second (FPS) out of your YOLOv5 model on edge devices, export your PyTorch weights to TensorRT for NVIDIA GPUs or [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) for Intel CPUs. This step can often double your inference speed. + To extract the maximum frames-per-second (FPS) out of your YOLOv5 model on edge devices, export your PyTorch weights to TensorRT for NVIDIA GPUs or [OpenVINO](https://docs.ultralytics.com/integrations/openvino) for Intel CPUs. This step can often double your inference speed. ## Training Ecosystem and Developer Experience The true advantage of the Ultralytics ecosystem lies in its streamlined user experience. While EfficientDet requires deep knowledge of the TensorFlow object detection API, YOLOv5 provides a consistent, simple Python API. -The well-maintained [Ultralytics ecosystem](https://docs.ultralytics.com/integrations/) ensures developers have access to frequent updates, active community support, and seamless integrations with experiment tracking tools like Weights & Biases and ClearML. +The well-maintained [Ultralytics ecosystem](https://docs.ultralytics.com/integrations) ensures developers have access to frequent updates, active community support, and seamless integrations with experiment tracking tools like Weights & Biases and ClearML. ### Code Example: Getting Started with YOLOv5 @@ -111,7 +111,7 @@ results[0].show() ## Versatility and Real-World Applications -EfficientDet is strictly an object detection framework, which limits its utility in complex vision pipelines. On the other hand, YOLOv5 has evolved to support multiple computer vision tasks. Modern releases of the model support highly accurate [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [image classification](https://docs.ultralytics.com/tasks/classify/), allowing developers to consolidate their machine learning stack. +EfficientDet is strictly an object detection framework, which limits its utility in complex vision pipelines. On the other hand, YOLOv5 has evolved to support multiple computer vision tasks. Modern releases of the model support highly accurate [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [image classification](https://docs.ultralytics.com/tasks/classify), allowing developers to consolidate their machine learning stack. ### Ideal Use Cases @@ -132,7 +132,7 @@ YOLO26 redefines the Pareto frontier of speed and accuracy, introducing groundbr - **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, which is critical for aerial imagery and robotics. - **DFL Removal:** By stripping out Distribution Focal Loss, the model head is greatly simplified, leading to better compatibility when exporting to legacy or highly constrained edge hardware. -For teams deploying multi-task pipelines, YOLO26 also introduces task-specific upgrades, such as multi-scale proto for segmentation and specialized angle loss for [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). To explore other modern alternatives within the ecosystem, you can also review [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or the YOLOv8 architecture. +For teams deploying multi-task pipelines, YOLO26 also introduces task-specific upgrades, such as multi-scale proto for segmentation and specialized angle loss for [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb). To explore other modern alternatives within the ecosystem, you can also review [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or the YOLOv8 architecture. ## Conclusion diff --git a/docs/en/compare/efficientdet-vs-yolov6.md b/docs/en/compare/efficientdet-vs-yolov6.md index d61ad733e8f..e1487132a4f 100644 --- a/docs/en/compare/efficientdet-vs-yolov6.md +++ b/docs/en/compare/efficientdet-vs-yolov6.md @@ -6,7 +6,7 @@ keywords: EfficientDet, YOLOv6, object detection, computer vision, model compari # EfficientDet vs YOLOv6-3.0: A Comprehensive Guide to Industrial Object Detection -Choosing the right neural network architecture is the cornerstone of any successful [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) initiative. This deep dive provides a highly technical comparison between two pivotal models in the [object detection](https://docs.ultralytics.com/tasks/detect/) landscape: Google's EfficientDet and Meituan's YOLOv6-3.0. +Choosing the right neural network architecture is the cornerstone of any successful [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) initiative. This deep dive provides a highly technical comparison between two pivotal models in the [object detection](https://docs.ultralytics.com/tasks/detect) landscape: Google's EfficientDet and Meituan's YOLOv6-3.0. While both architectures represented major leaps forward upon their respective releases, the rapid evolution of artificial intelligence has introduced more versatile, edge-optimized solutions. Below, we dissect the performance, training methodologies, and architectural nuances of EfficientDet and YOLOv6-3.0, and explore why developers are increasingly migrating to modern ecosystems like [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) for state-of-the-art deployment. @@ -45,7 +45,7 @@ Released to serve the specific needs of bulk processing, YOLOv6-3.0 is a [convol - **Date:** 2023-01-13 - **Arxiv:** [2301.05586](https://arxiv.org/abs/2301.05586) - **GitHub:** [meituan/YOLOv6](https://github.com/meituan/YOLOv6) -- **Docs:** [YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6/) +- **Docs:** [YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6) ### Architectural Innovations @@ -55,11 +55,11 @@ YOLOv6-3.0 replaces traditional modules with the **Bi-directional Concatenation Built on the hardware-friendly EfficientRep backbone, YOLOv6-3.0 excels in high-speed industrial [manufacturing environments](https://www.ultralytics.com/solutions/ai-in-manufacturing) where batch processing on dedicated GPUs is possible. However, its heavy reliance on re-parameterization operations can lead to significant drops in speed when deployed on edge devices or environments relying strictly on CPU computations. -[Learn more about YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Performance Comparison -Understanding the raw [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/) is fundamental to selecting a model that aligns with your specific deployment constraints. Below is a detailed breakdown of accuracy, speed, and computational footprint. +Understanding the raw [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics) is fundamental to selecting a model that aligns with your specific deployment constraints. Below is a detailed breakdown of accuracy, speed, and computational footprint. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | --------------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -103,11 +103,11 @@ YOLOv6 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Why YOLO26 is the Superior Choice @@ -117,15 +117,15 @@ For developers seeking the absolute peak of performance and ease of use, **Ultra ### YOLO26 Breakthrough Innovations -- **End-to-End NMS-Free Design:** YOLO26 is natively end-to-end, completely eliminating the need for Non-Maximum Suppression (NMS) post-processing. This drastically reduces latency variance and simplifies [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/) across diverse edge hardware. +- **End-to-End NMS-Free Design:** YOLO26 is natively end-to-end, completely eliminating the need for Non-Maximum Suppression (NMS) post-processing. This drastically reduces latency variance and simplifies [model deployment](https://docs.ultralytics.com/guides/model-deployment-options) across diverse edge hardware. - **MuSGD Optimizer:** Inspired by LLM training (like Moonshot AI's Kimi K2), YOLO26 utilizes a hybrid of SGD and Muon. This brings large language model stability to computer vision, ensuring faster convergence and highly efficient training processes. - **Up to 43% Faster CPU Inference:** Optimized specifically for [edge computing](https://www.ultralytics.com/glossary/edge-computing) and low-power devices, YOLO26 delivers unmatched CPU speeds where traditional industrial models struggle. -- **DFL Removal:** The Distribution Focal Loss has been removed to simplify the export graph, granting seamless compatibility with deployment runtimes like [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) and CoreML. +- **DFL Removal:** The Distribution Focal Loss has been removed to simplify the export graph, granting seamless compatibility with deployment runtimes like [OpenVINO](https://docs.ultralytics.com/integrations/openvino) and CoreML. - **ProgLoss + STAL:** Advanced loss functions provide notable improvements in [small-object recognition](https://www.ultralytics.com/blog/exploring-small-object-detection-with-ultralytics-yolo11), making YOLO26 indispensable for drone mapping, IoT sensors, and robotics. ### Unmatched Versatility -Unlike EfficientDet, which is confined to bounding box detection, YOLO26 is a natively multi-task learner. The same unified [Python API](https://docs.ultralytics.com/usage/python/) supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), Image Classification, and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection out of the box, with task-specific improvements like Semantic Segmentation Loss and Residual Log-Likelihood Estimation (RLE) built directly into the architecture. +Unlike EfficientDet, which is confined to bounding box detection, YOLO26 is a natively multi-task learner. The same unified [Python API](https://docs.ultralytics.com/usage/python) supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), Image Classification, and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection out of the box, with task-specific improvements like Semantic Segmentation Loss and Residual Log-Likelihood Estimation (RLE) built directly into the architecture. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -155,4 +155,4 @@ model.export(format="onnx") If your project requires supporting older hardware profiles or you are maintaining a legacy codebase, the broader Ultralytics ecosystem has you covered. - **[Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11):** The immediate predecessor to YOLO26, highly trusted in enterprise environments requiring mature, well-documented pipelines. -- **[Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8):** The standard-bearer that redefined the developer experience, remaining an excellent choice for general-purpose computer vision tasks integrated deeply with tools like [TensorBoard](https://docs.ultralytics.com/integrations/tensorboard/) and [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/). +- **[Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8):** The standard-bearer that redefined the developer experience, remaining an excellent choice for general-purpose computer vision tasks integrated deeply with tools like [TensorBoard](https://docs.ultralytics.com/integrations/tensorboard) and [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases). diff --git a/docs/en/compare/efficientdet-vs-yolov7.md b/docs/en/compare/efficientdet-vs-yolov7.md index 6a3958fd86a..13f7c7959a5 100644 --- a/docs/en/compare/efficientdet-vs-yolov7.md +++ b/docs/en/compare/efficientdet-vs-yolov7.md @@ -8,7 +8,7 @@ keywords: EfficientDet, YOLOv7, object detection, model comparison, EfficientDet Selecting the most effective neural network architecture is critical to the success of any [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) initiative. As the demand for high-performance AI solutions accelerates, comparing established models like EfficientDet and YOLOv7 becomes essential for developers aiming to optimize both accuracy and computational efficiency. -This comprehensive technical analysis explores the architectural nuances, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/), and ideal deployment scenarios for both models. Additionally, we will illustrate why the integrated ecosystem provided by Ultralytics—culminating in the state-of-the-art [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26)—offers a superior alternative for modern computer vision tasks. +This comprehensive technical analysis explores the architectural nuances, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics), and ideal deployment scenarios for both models. Additionally, we will illustrate why the integrated ecosystem provided by Ultralytics—culminating in the state-of-the-art [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26)—offers a superior alternative for modern computer vision tasks. @@ -59,7 +59,7 @@ YOLOv7 excels in real-time scenarios, such as [video analytics](https://www.ultr Despite its impressive speed, YOLOv7 still relies on Non-Maximum Suppression (NMS) for post-processing, which can introduce variable latency in crowded scenes. Furthermore, its memory footprint during training is notably larger than newer generations, requiring more robust hardware to handle large batch sizes. -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ## Performance and Metrics Comparison @@ -105,20 +105,20 @@ YOLOv7 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage Choosing the right architecture goes beyond just raw metrics; it involves evaluating the entire machine learning lifecycle. The [Ultralytics ecosystem](https://docs.ultralytics.com/) provides an unparalleled developer experience, significantly lowering the barrier to entry for robust AI deployments. - **Ease of Use:** Ultralytics provides a highly unified Python API. Developers can train, validate, and export models in just a few lines of code, removing the need to manage complex, fragmented codebases typical of EfficientDet. -- **Well-Maintained Ecosystem:** Benefiting from rapid updates, extensive documentation, and an active community, Ultralytics ensures compatibility with the latest [deployment frameworks](https://docs.ultralytics.com/guides/model-deployment-options/) like TensorRT and OpenVINO. +- **Well-Maintained Ecosystem:** Benefiting from rapid updates, extensive documentation, and an active community, Ultralytics ensures compatibility with the latest [deployment frameworks](https://docs.ultralytics.com/guides/model-deployment-options) like TensorRT and OpenVINO. - **Memory Requirements:** By utilizing highly optimized PyTorch data loaders and streamlined network structures, Ultralytics YOLO models require significantly less CUDA memory during training compared to multi-branch networks and transformer-heavy models. -- **Versatility:** Unlike older architectures strictly tied to bounding box detection, Ultralytics models are multi-task powerhouses supporting [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +- **Versatility:** Unlike older architectures strictly tied to bounding box detection, Ultralytics models are multi-task powerhouses supporting [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). ### Training Efficiency with Ultralytics @@ -143,11 +143,11 @@ While YOLOv7 and EfficientDet laid the groundwork for modern computer vision, th ### Key YOLO26 Innovations -- **End-to-End NMS-Free Design:** Building on the foundations laid by [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 is natively end-to-end. By entirely eliminating Non-Maximum Suppression (NMS) post-processing, it delivers lower, more consistent latency, which is crucial for safety-critical systems like autonomous driving. +- **End-to-End NMS-Free Design:** Building on the foundations laid by [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 is natively end-to-end. By entirely eliminating Non-Maximum Suppression (NMS) post-processing, it delivers lower, more consistent latency, which is crucial for safety-critical systems like autonomous driving. - **Up to 43% Faster CPU Inference:** Thanks to the **removal of Distribution Focal Loss (DFL)**, YOLO26 features a drastically simplified export process and unparalleled speed on edge devices like the Raspberry Pi, making it the undisputed champion of edge computing. - **MuSGD Optimizer:** YOLO26 incorporates the revolutionary MuSGD Optimizer—a hybrid of SGD and Muon inspired by LLM training innovations from Moonshot AI. This leads to remarkably stable training dynamics and much faster convergence rates. -- **ProgLoss + STAL:** The integration of Progressive Loss and Scale-Targeted Alignment Loss heavily improves the model's ability to detect tiny objects, solving a massive pain point for drone imagery and [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/). -- **Task-Specific Improvements:** YOLO26 isn't just a detector. It features a Semantic segmentation loss and multi-scale proto for flawless [segmentation](https://docs.ultralytics.com/tasks/segment/), Residual Log-Likelihood Estimation (RLE) for hyper-accurate [pose tracking](https://docs.ultralytics.com/tasks/pose/), and specialized angle loss for resolving [OBB](https://docs.ultralytics.com/tasks/obb/) boundary ambiguities. +- **ProgLoss + STAL:** The integration of Progressive Loss and Scale-Targeted Alignment Loss heavily improves the model's ability to detect tiny objects, solving a massive pain point for drone imagery and [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system). +- **Task-Specific Improvements:** YOLO26 isn't just a detector. It features a Semantic segmentation loss and multi-scale proto for flawless [segmentation](https://docs.ultralytics.com/tasks/segment), Residual Log-Likelihood Estimation (RLE) for hyper-accurate [pose tracking](https://docs.ultralytics.com/tasks/pose), and specialized angle loss for resolving [OBB](https://docs.ultralytics.com/tasks/obb) boundary ambiguities. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -155,6 +155,6 @@ While YOLOv7 and EfficientDet laid the groundwork for modern computer vision, th While YOLO26 represents the pinnacle of current technology, the Ultralytics ecosystem supports a variety of models tailored for different use cases. -For developers managing legacy systems that still require traditional anchor-free scaling, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a robust, highly supported option within the Ultralytics platform. Additionally, for scenarios explicitly demanding transformer-based architectures, [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) offers real-time detection utilizing vision transformers, bridging the gap between high-end attention mechanisms and real-time execution speeds. +For developers managing legacy systems that still require traditional anchor-free scaling, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a robust, highly supported option within the Ultralytics platform. Additionally, for scenarios explicitly demanding transformer-based architectures, [RT-DETR](https://docs.ultralytics.com/models/rtdetr) offers real-time detection utilizing vision transformers, bridging the gap between high-end attention mechanisms and real-time execution speeds. -In conclusion, while EfficientDet provides academic insights into compound scaling and YOLOv7 offers strong baseline real-time performance, modern enterprises are best served by adopting the [Ultralytics Platform](https://docs.ultralytics.com/platform/). By leveraging YOLO26, teams can ensure maximum performance, minimal training friction, and future-proof their AI deployments. +In conclusion, while EfficientDet provides academic insights into compound scaling and YOLOv7 offers strong baseline real-time performance, modern enterprises are best served by adopting the [Ultralytics Platform](https://docs.ultralytics.com/platform). By leveraging YOLO26, teams can ensure maximum performance, minimal training friction, and future-proof their AI deployments. diff --git a/docs/en/compare/efficientdet-vs-yolov8.md b/docs/en/compare/efficientdet-vs-yolov8.md index 5c45ce738b9..2c7633b2eff 100644 --- a/docs/en/compare/efficientdet-vs-yolov8.md +++ b/docs/en/compare/efficientdet-vs-yolov8.md @@ -48,7 +48,7 @@ YOLOv8 introduced an **anchor-free** detection head, eliminating the need to man !!! tip "Multi-Task Capabilities" - Unlike EfficientDet, which is strictly designed for bounding boxes, YOLOv8 boasts extreme **versatility**. Out of the box, it supports [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). + Unlike EfficientDet, which is strictly designed for bounding boxes, YOLOv8 boasts extreme **versatility**. Out of the box, it supports [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb). ## Performance and Benchmarks @@ -83,7 +83,7 @@ The true differentiator for many modern engineering teams is not just the raw sp EfficientDet's implementation relies heavily on legacy AutoML libraries, which can present a steep learning curve and brittle dependency chains for developers accustomed to modern [PyTorch](https://pytorch.org/) workflows. -In contrast, Ultralytics offers an unparalleled **ease of use**. The [well-maintained ecosystem](https://docs.ultralytics.com/) provides a consistent Python API that drastically simplifies the machine learning lifecycle. It offers seamless integration with the robust [Ultralytics Platform](https://docs.ultralytics.com/platform/), which handles everything from auto-annotation to cloud training and real-time monitoring. +In contrast, Ultralytics offers an unparalleled **ease of use**. The [well-maintained ecosystem](https://docs.ultralytics.com/) provides a consistent Python API that drastically simplifies the machine learning lifecycle. It offers seamless integration with the robust [Ultralytics Platform](https://docs.ultralytics.com/platform), which handles everything from auto-annotation to cloud training and real-time monitoring. ### Code Example: Training and Inference with YOLOv8 @@ -105,7 +105,7 @@ predictions = model("https://ultralytics.com/images/bus.jpg") export_path = model.export(format="onnx") ``` -This streamlined approach automatically handles dataset downloading, [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation/), and hardware allocation, allowing researchers to focus on results rather than boilerplate code. +This streamlined approach automatically handles dataset downloading, [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation), and hardware allocation, allowing researchers to focus on results rather than boilerplate code. ## Use Cases and Recommendations @@ -123,23 +123,23 @@ EfficientDet is a strong choice for: YOLOv8 is recommended for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Looking Forward: The YOLO26 Advantage While YOLOv8 is a fantastic general-purpose model, the computer vision landscape has continued to advance. For users evaluating architectures today, it is highly recommended to explore the newly released [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26), which represents the pinnacle of modern object detection. -Released in January 2026, YOLO26 builds upon the successes of its predecessors (including [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv10](https://docs.ultralytics.com/models/yolov10/)) with groundbreaking features: +Released in January 2026, YOLO26 builds upon the successes of its predecessors (including [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv10](https://docs.ultralytics.com/models/yolov10)) with groundbreaking features: - **End-to-End NMS-Free Design:** YOLO26 natively eliminates the need for Non-Maximum Suppression (NMS) post-processing, vastly simplifying deployment logic and reducing latency variance. - **MuSGD Optimizer:** Integrating innovations from Large Language Model (LLM) training, this hybrid optimizer ensures more stable training and rapid convergence. diff --git a/docs/en/compare/efficientdet-vs-yolov9.md b/docs/en/compare/efficientdet-vs-yolov9.md index fc064e1c2e7..f827a30d106 100644 --- a/docs/en/compare/efficientdet-vs-yolov9.md +++ b/docs/en/compare/efficientdet-vs-yolov9.md @@ -6,7 +6,7 @@ keywords: EfficientDet, YOLOv9, object detection comparison, computer vision, mo # EfficientDet vs. YOLOv9: Architecture, Performance, and Edge Deployment -The landscape of computer vision has been shaped by continuous breakthroughs in neural network design. Finding the right balance between computational efficiency and detection accuracy is critical when selecting a model. Google's **EfficientDet** established a strong baseline in 2019 by introducing scalable architectures, while **YOLOv9**, released in 2024, pushed the boundaries of [object detection](https://docs.ultralytics.com/tasks/detect/) using Programmable Gradient Information (PGI). +The landscape of computer vision has been shaped by continuous breakthroughs in neural network design. Finding the right balance between computational efficiency and detection accuracy is critical when selecting a model. Google's **EfficientDet** established a strong baseline in 2019 by introducing scalable architectures, while **YOLOv9**, released in 2024, pushed the boundaries of [object detection](https://docs.ultralytics.com/tasks/detect) using Programmable Gradient Information (PGI). This guide provides a comprehensive technical comparison between these two models and introduces the modern [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) framework, which offers a robust, end-to-end solution optimized for production environments. @@ -40,16 +40,16 @@ Developed by researchers at Academia Sinica, YOLOv9 tackles the degradation of i - **Authors:** Chien-Yao Wang and Hong-Yuan Mark Liao - **Organization:** Institute of Information Science, Academia Sinica - **Date:** February 21, 2024 -- **Links:** [Arxiv](https://arxiv.org/abs/2402.13616), [GitHub](https://github.com/WongKinYiu/yolov9), [Docs](https://docs.ultralytics.com/models/yolov9/) +- **Links:** [Arxiv](https://arxiv.org/abs/2402.13616), [GitHub](https://github.com/WongKinYiu/yolov9), [Docs](https://docs.ultralytics.com/models/yolov9) **Key Architectural Features:** YOLOv9 introduces Programmable Gradient Information (PGI) to provide auxiliary supervision, ensuring crucial data is retained for updating network weights reliably. It also features the Generalized Efficient Layer Aggregation Network (GELAN) to maximize parameter efficiency. Despite these advancements, YOLOv9 still requires Non-Maximum Suppression (NMS) during post-processing, which adds latency. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ## Performance Comparison -When evaluating these models, analyzing empirical data helps determine which architecture provides the best trade-off for your specific [hardware requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics/). +When evaluating these models, analyzing empirical data helps determine which architecture provides the best trade-off for your specific [hardware requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | --------------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -74,7 +74,7 @@ YOLOv9 provides a generational leap in speed. For instance, YOLOv9e achieves a * !!! tip "Exporting Models for Production" - Exporting your architecture to optimized formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) drastically reduces inference times compared to raw PyTorch runs. + Exporting your architecture to optimized formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino) drastically reduces inference times compared to raw PyTorch runs. ## Use Cases and Recommendations @@ -98,11 +98,11 @@ YOLOv9 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Choosing YOLO26 @@ -121,11 +121,11 @@ The [Ultralytics Platform](https://platform.ultralytics.com) offers unparalleled [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } -Other robust options in the Ultralytics ecosystem include [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), which also provide multi-task versatility such as [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [pose estimation](https://docs.ultralytics.com/tasks/pose/). +Other robust options in the Ultralytics ecosystem include [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), which also provide multi-task versatility such as [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [pose estimation](https://docs.ultralytics.com/tasks/pose). ### Simplified Training with the Python SDK -Ultralytics models prioritize developer experience. Training a state-of-the-art model is condensed into just a few lines of [Python](https://docs.ultralytics.com/usage/python/). +Ultralytics models prioritize developer experience. Training a state-of-the-art model is condensed into just a few lines of [Python](https://docs.ultralytics.com/usage/python). ```python from ultralytics import YOLO @@ -147,4 +147,4 @@ Choosing between these architectures heavily depends on your deployment target. - **Legacy Cloud Deployments:** EfficientDet was popular for offline, cloud-based batch processing where high accuracy was needed, and strict real-time constraints were non-existent. - **Academic Research:** YOLOv9 remains an interesting choice for researchers pushing theoretical CNN bounds and analyzing gradient flows through network layers. -- **Edge Computing and IoT:** **YOLO26** dominates real-world applications. Its NMS-free pipeline and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) capabilities make it the superior option for smart city traffic analysis, retail inventory monitoring, and drone-based inspection, offering an unbeatable balance of high accuracy and rapid inference speeds. +- **Edge Computing and IoT:** **YOLO26** dominates real-world applications. Its NMS-free pipeline and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) capabilities make it the superior option for smart city traffic analysis, retail inventory monitoring, and drone-based inspection, offering an unbeatable balance of high accuracy and rapid inference speeds. diff --git a/docs/en/compare/efficientdet-vs-yolox.md b/docs/en/compare/efficientdet-vs-yolox.md index 8b1345e8e7e..721450b5f3b 100644 --- a/docs/en/compare/efficientdet-vs-yolox.md +++ b/docs/en/compare/efficientdet-vs-yolox.md @@ -51,13 +51,13 @@ Released two years later, YOLOX sought to bridge the gap between academic resear YOLOX significantly simplified the object detection paradigm. By switching to an **anchor-free** design, YOLOX eliminated the need for complex, dataset-specific anchor box tuning, reducing heuristic overhead. It also integrated a decoupled head—separating classification and localization tasks—which drastically improved convergence speed. Furthermore, the introduction of the **SimOTA** label assignment strategy optimized the allocation of positive samples dynamically during training. -Despite these advancements, managing YOLOX repositories often requires compiling manual C++ extensions and navigating complex dependencies, which can hinder rapid [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/) for less experienced teams. +Despite these advancements, managing YOLOX repositories often requires compiling manual C++ extensions and navigating complex dependencies, which can hinder rapid [model deployment](https://docs.ultralytics.com/guides/model-deployment-options) for less experienced teams. [Learn more about YOLOX](https://yolox.readthedocs.io/en/latest/){ .md-button } ## Performance Comparison -When evaluating models for production, balancing [mean Average Precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics/) with inference speed is paramount. The table below provides a direct comparison of the EfficientDet and YOLOX families across standard COCO benchmarks. +When evaluating models for production, balancing [mean Average Precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics) with inference speed is paramount. The table below provides a direct comparison of the EfficientDet and YOLOX families across standard COCO benchmarks. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | --------------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -103,15 +103,15 @@ YOLOX is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Introducing YOLO26 -While EfficientDet and YOLOX represented significant leaps in their respective eras, modern computer vision demands greater versatility, streamlined workflows, and uncompromising speed. For developers prioritizing ease of use, lower memory requirements, and a well-maintained ecosystem, we highly recommend upgrading to **[Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/)**, released in January 2026. +While EfficientDet and YOLOX represented significant leaps in their respective eras, modern computer vision demands greater versatility, streamlined workflows, and uncompromising speed. For developers prioritizing ease of use, lower memory requirements, and a well-maintained ecosystem, we highly recommend upgrading to **[Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26)**, released in January 2026. YOLO26 represents a paradigm shift in the YOLO lineage, systematically overcoming the limitations found in older models like YOLOX and EfficientDet: @@ -119,13 +119,13 @@ YOLO26 represents a paradigm shift in the YOLO lineage, systematically overcomin - **Up to 43% Faster CPU Inference:** Through strategic architectural tuning and the **DFL Removal** (Distribution Focal Loss), YOLO26 is uniquely optimized for environments without dedicated GPUs, completely outpacing EfficientDet on [edge AI](https://www.ultralytics.com/glossary/edge-ai) hardware like Raspberry Pi. - **MuSGD Optimizer:** Inspired by LLM training innovations (like Moonshot AI's Kimi K2), YOLO26 uses a hybrid of SGD and Muon. This ensures incredibly stable training and faster convergence, vastly superior to older TensorFlow estimators. - **ProgLoss + STAL:** Advanced loss functions bring notable improvements in small-object recognition, a historic weakness for both YOLOX and EfficientDet. This is critical for drone analytics and IoT. -- **Incredible Versatility:** While EfficientDet and YOLOX are strictly bounding box detectors, YOLO26 natively supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/) (via Residual Log-Likelihood Estimation), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +- **Incredible Versatility:** While EfficientDet and YOLOX are strictly bounding box detectors, YOLO26 natively supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose) (via Residual Log-Likelihood Estimation), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } ### Streamlined User Experience and Training Efficiency -One of the largest hurdles with models like YOLOX is setting up the training environment. The [Ultralytics Platform](https://platform.ultralytics.com/) offers a unified [Python SDK](https://docs.ultralytics.com/usage/python/) where training a state-of-the-art model requires only a few lines of code. Additionally, YOLO models feature highly optimized data loaders, ensuring significantly lower CUDA memory usage compared to transformer-heavy models or older multi-branch networks. +One of the largest hurdles with models like YOLOX is setting up the training environment. The [Ultralytics Platform](https://platform.ultralytics.com/) offers a unified [Python SDK](https://docs.ultralytics.com/usage/python) where training a state-of-the-art model requires only a few lines of code. Additionally, YOLO models feature highly optimized data loaders, ensuring significantly lower CUDA memory usage compared to transformer-heavy models or older multi-branch networks. ```python from ultralytics import YOLO diff --git a/docs/en/compare/index.md b/docs/en/compare/index.md index a2fbfa1da43..c56259cbcf2 100644 --- a/docs/en/compare/index.md +++ b/docs/en/compare/index.md @@ -6,20 +6,20 @@ keywords: YOLO26, YOLO11, YOLOv10 comparison, YOLOv8 vs RT-DETR, object detectio # Model Comparisons: Choose the Best Object Detection Model for Your Project -Choosing the right neural network architecture is the cornerstone of any successful [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) project. Welcome to the **Ultralytics Model Comparison Hub**! This page centralizes detailed technical analyses and performance benchmarks, dissecting the trade-offs between the latest [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) and other leading architectures like YOLO11, YOLOv10, RT-DETR, and EfficientDet. +Choosing the right neural network architecture is the cornerstone of any successful [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) project. Welcome to the **Ultralytics Model Comparison Hub**! This page centralizes detailed technical analyses and performance benchmarks, dissecting the trade-offs between the latest [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) and other leading architectures like YOLO11, YOLOv10, RT-DETR, and EfficientDet. Whether your application demands the millisecond latency of [edge AI](https://www.ultralytics.com/glossary/edge-ai) or the high-fidelity precision required for medical imaging, this guide provides the data-driven insights needed to make an informed choice. We evaluate models based on [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map), [inference speed](https://www.ultralytics.com/glossary/inference-latency), parameter efficiency, and ease of deployment. ## Interactive Performance Benchmarks -Visualizing the relationship between speed and accuracy is essential for identifying the "Pareto frontier" of object detection—models that offer the best accuracy for a given speed constraint. The chart below contrasts key metrics on standard [datasets like COCO](https://docs.ultralytics.com/datasets/detect/coco/). +Visualizing the relationship between speed and accuracy is essential for identifying the "Pareto frontier" of object detection—models that offer the best accuracy for a given speed constraint. The chart below contrasts key metrics on standard [datasets like COCO](https://docs.ultralytics.com/datasets/detect/coco). -This chart visualizes key [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/) enabling you to quickly assess the trade-offs between different models. Understanding these metrics is fundamental to selecting a model that aligns with your specific deployment constraints. +This chart visualizes key [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics) enabling you to quickly assess the trade-offs between different models. Understanding these metrics is fundamental to selecting a model that aligns with your specific deployment constraints. ## Quick Decision Guide diff --git a/docs/en/compare/pp-yoloe-vs-damo-yolo.md b/docs/en/compare/pp-yoloe-vs-damo-yolo.md index 829c24fde89..87ae7bce419 100644 --- a/docs/en/compare/pp-yoloe-vs-damo-yolo.md +++ b/docs/en/compare/pp-yoloe-vs-damo-yolo.md @@ -99,7 +99,7 @@ Released in early 2026, YOLO26 builds upon the legacy of [YOLO11](https://platfo - **MuSGD Optimizer:** Inspired by large language model training techniques, YOLO26 utilizes a hybrid MuSGD optimizer. This ensures incredibly stable training and rapid convergence, saving valuable GPU hours. - **Superior CPU Inference:** By removing Distribution Focal Loss (DFL) and optimizing the network graph, YOLO26 achieves up to 43% faster CPU inference, making it the premier choice for [edge AI devices](https://www.ultralytics.com/glossary/edge-ai). - **ProgLoss + STAL:** These advanced loss functions yield remarkable improvements in small-object recognition, which is critical for [drone operations](https://www.ultralytics.com/solutions/ai-in-agriculture) and remote sensing. -- **Unmatched Versatility:** Unlike PP-YOLOE+ which focuses strictly on detection, YOLO26 natively supports [pose estimation](https://docs.ultralytics.com/tasks/pose/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) seamlessly. +- **Unmatched Versatility:** Unlike PP-YOLOE+ which focuses strictly on detection, YOLO26 natively supports [pose estimation](https://docs.ultralytics.com/tasks/pose), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb) seamlessly. ### Ease of Use and Training Efficiency @@ -131,4 +131,4 @@ Selecting the optimal computer vision architecture depends heavily on your team' - **Choose DAMO-YOLO** if you are conducting specific research into Neural Architecture Search algorithms, or if you have the engineering resources to maintain complex distillation pipelines to achieve aggressive TensorRT latency targets. - **Choose Ultralytics YOLO26** for almost all modern production scenarios. The [Ultralytics ecosystem](https://www.ultralytics.com/) provides unparalleled documentation, lower memory requirements, and a streamlined API. Whether you are building [automated quality control](https://www.ultralytics.com/solutions/ai-in-manufacturing) systems or running real-time tracking on a Raspberry Pi, YOLO26’s NMS-free architecture ensures rapid, stable, and highly accurate results out of the box. -For developers exploring other state-of-the-art solutions, the Ultralytics documentation also provides extensive resources on the widely adopted [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) and the robust [YOLO11](https://docs.ultralytics.com/models/yolo11/), ensuring you have the right model for any computer vision challenge. +For developers exploring other state-of-the-art solutions, the Ultralytics documentation also provides extensive resources on the widely adopted [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) and the robust [YOLO11](https://docs.ultralytics.com/models/yolo11), ensuring you have the right model for any computer vision challenge. diff --git a/docs/en/compare/pp-yoloe-vs-efficientdet.md b/docs/en/compare/pp-yoloe-vs-efficientdet.md index b0ddf53a16a..a29d3f87d64 100644 --- a/docs/en/compare/pp-yoloe-vs-efficientdet.md +++ b/docs/en/compare/pp-yoloe-vs-efficientdet.md @@ -6,7 +6,7 @@ keywords: PP-YOLOE+,EfficientDet,object detection,PP-YOLOE+m,EfficientDet-D7,AI # PP-YOLOE+ vs EfficientDet: A Comprehensive Technical Comparison -Choosing the right architecture is a critical step in building robust [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) applications. This technical guide explores the trade-offs between two well-known object detection models: **PP-YOLOE+** and **EfficientDet**. We will break down their architectures, analyze their [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/), and explore their ideal deployment scenarios. +Choosing the right architecture is a critical step in building robust [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) applications. This technical guide explores the trade-offs between two well-known object detection models: **PP-YOLOE+** and **EfficientDet**. We will break down their architectures, analyze their [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics), and explore their ideal deployment scenarios. While both models have made significant contributions to the field, we will also discuss how modern alternatives like [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) offer vastly superior memory efficiency, faster inference, and a highly streamlined developer experience. @@ -31,11 +31,11 @@ PP-YOLOE+ features a CSPRepResNet backbone, an Efficient Task-aligned head (ET-h !!! tip "Integration Benefits" - Teams already deeply invested in Baidu's PaddlePaddle framework often find PP-YOLOE+ easier to adopt for tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment/), though it lacks the broad multi-framework support seen in newer tools. + Teams already deeply invested in Baidu's PaddlePaddle framework often find PP-YOLOE+ easier to adopt for tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment), though it lacks the broad multi-framework support seen in newer tools. ## Architectural Overview: EfficientDet -EfficientDet takes a radically different approach to [object detection](https://docs.ultralytics.com/tasks/detect/), relying heavily on neural architecture search and compound scaling principles. +EfficientDet takes a radically different approach to [object detection](https://docs.ultralytics.com/tasks/detect), relying heavily on neural architecture search and compound scaling principles. - **Authors:** Mingxing Tan, Ruoming Pang, and Quoc V. Le - **Organization:** [Google](https://github.com/google/automl/tree/master/efficientdet) @@ -51,7 +51,7 @@ While theoretically highly efficient in terms of FLOPs, EfficientDet models can ## Performance Analysis and Benchmarks -The table below contrasts key metrics on standard [datasets like COCO](https://docs.ultralytics.com/datasets/detect/coco/). Comparing [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map) against inference speed provides a clear picture of the Pareto frontier. +The table below contrasts key metrics on standard [datasets like COCO](https://docs.ultralytics.com/datasets/detect/coco). Comparing [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map) against inference speed provides a clear picture of the Pareto frontier. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | --------------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -94,11 +94,11 @@ EfficientDet is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Modern Alternative: Ultralytics YOLO26 @@ -106,14 +106,14 @@ While PP-YOLOE+ and EfficientDet represent significant historical milestones, de YOLO26 represents a massive leap forward in object detection, introducing several critical innovations: -- **End-to-End NMS-Free Design:** Building on the breakthroughs of [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates Non-Maximum Suppression (NMS) during inference. This results in significantly lower latency and removes complex post-processing bottlenecks. +- **End-to-End NMS-Free Design:** Building on the breakthroughs of [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates Non-Maximum Suppression (NMS) during inference. This results in significantly lower latency and removes complex post-processing bottlenecks. - **MuSGD Optimizer:** Inspired by LLM training innovations, YOLO26 utilizes a hybrid SGD and Muon optimizer. This drastically improves training stability and reduces convergence time. -- **Extreme Speed:** YOLO26 delivers up to **43% faster CPU inference** compared to older generations like [YOLO11](https://docs.ultralytics.com/models/yolo11/), making it the absolute best choice for battery-powered or CPU-only edge devices. +- **Extreme Speed:** YOLO26 delivers up to **43% faster CPU inference** compared to older generations like [YOLO11](https://docs.ultralytics.com/models/yolo11), making it the absolute best choice for battery-powered or CPU-only edge devices. - **Advanced Loss Functions:** The integration of ProgLoss and STAL greatly improves small-object recognition, which is essential for tasks like [drone analytics](https://www.ultralytics.com/blog/build-ai-powered-drone-applications-with-ultralytics-yolo11) and [robotics](https://www.ultralytics.com/blog/from-algorithms-to-automation-ais-role-in-robotics). !!! note "Multi-Task Versatility" - Unlike EfficientDet which focuses purely on detection, YOLO26 natively handles [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), all within the same well-maintained ecosystem. + Unlike EfficientDet which focuses purely on detection, YOLO26 natively handles [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb), all within the same well-maintained ecosystem. ### Ease of Use and Ecosystem Integration @@ -134,7 +134,7 @@ results = model.train(data="coco8.yaml", epochs=100) predictions = model("https://ultralytics.com/images/bus.jpg") ``` -For those exploring other alternatives, architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) or the legacy [YOLOv8](https://docs.ultralytics.com/models/yolov8/) are also available within the Ultralytics ecosystem, allowing for seamless swapping and testing. +For those exploring other alternatives, architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) or the legacy [YOLOv8](https://docs.ultralytics.com/models/yolov8) are also available within the Ultralytics ecosystem, allowing for seamless swapping and testing. ## Conclusion diff --git a/docs/en/compare/pp-yoloe-vs-rtdetr.md b/docs/en/compare/pp-yoloe-vs-rtdetr.md index 799cf3ce206..31702a2c7bd 100644 --- a/docs/en/compare/pp-yoloe-vs-rtdetr.md +++ b/docs/en/compare/pp-yoloe-vs-rtdetr.md @@ -61,7 +61,7 @@ RTDETRv2 leverages a hybrid architecture, combining a CNN backbone for feature e The transformer architecture makes RTDETRv2 highly effective in scenarios where understanding global context is crucial. However, transformer models typically demand significantly higher CUDA memory during both training and inference compared to lightweight CNNs. It is best suited for environments with unconstrained hardware, such as cloud-based [video analytics](https://www.ultralytics.com/blog/behind-the-scenes-of-vision-ai-in-streaming) running on powerful GPU servers. -[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr){ .md-button } ## Performance and Metrics Comparison @@ -104,11 +104,11 @@ RT-DETR is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Introducing YOLO26 @@ -130,7 +130,7 @@ YOLO26 introduces several pioneering enhancements that outclass traditional CNNs !!! note "Task-Specific Versatility" - Unlike specialized object detectors, YOLO26 is highly versatile, supporting [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). It includes tailored enhancements like RLE for Pose and specialized angle loss for OBB. + Unlike specialized object detectors, YOLO26 is highly versatile, supporting [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). It includes tailored enhancements like RLE for Pose and specialized angle loss for OBB. ### Unmatched Ease of Use @@ -157,7 +157,7 @@ model_yolo.export(format="engine", half=True) Lower memory requirements typical of Ultralytics YOLO models mean you can train faster and deploy on cheaper hardware compared to transformer-based counterparts. Furthermore, active development and world-class documentation ensure your production pipelines remain stable. -For teams exploring alternatives, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a highly supported and exceptionally capable predecessor within the ecosystem, providing an excellent baseline for legacy hardware integrations. You might also find it useful to read our comparison on [YOLO11 vs RTDETR](https://docs.ultralytics.com/compare/yolo11-vs-rtdetr/). +For teams exploring alternatives, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a highly supported and exceptionally capable predecessor within the ecosystem, providing an excellent baseline for legacy hardware integrations. You might also find it useful to read our comparison on [YOLO11 vs RTDETR](https://docs.ultralytics.com/compare/yolo11-vs-rtdetr). ## Summary diff --git a/docs/en/compare/pp-yoloe-vs-yolo11.md b/docs/en/compare/pp-yoloe-vs-yolo11.md index 299975f4a6e..063a98e2668 100644 --- a/docs/en/compare/pp-yoloe-vs-yolo11.md +++ b/docs/en/compare/pp-yoloe-vs-yolo11.md @@ -6,7 +6,7 @@ keywords: PP-YOLOE+, YOLO11, object detection, model comparison, computer vision # A Deep Dive into Real-Time Object Detection: PP-YOLOE+ vs YOLO11 -The landscape of computer vision is constantly evolving, driven by the need for faster, more accurate, and more efficient models. For developers and researchers tackling [object detection](https://docs.ultralytics.com/tasks/detect/) tasks, choosing the right architecture is critical. In this comprehensive comparison, we will explore the nuances between two prominent models: **PP-YOLOE+** and **Ultralytics YOLO11**. +The landscape of computer vision is constantly evolving, driven by the need for faster, more accurate, and more efficient models. For developers and researchers tackling [object detection](https://docs.ultralytics.com/tasks/detect) tasks, choosing the right architecture is critical. In this comprehensive comparison, we will explore the nuances between two prominent models: **PP-YOLOE+** and **Ultralytics YOLO11**. By dissecting their architectures, performance metrics, and ideal use cases, this guide aims to provide the insights necessary to make an informed decision for your next machine learning deployment. @@ -40,25 +40,25 @@ YOLO11, created by Ultralytics, represents a significant leap forward in usabili - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2024-09-27 - **GitHub:** [Ultralytics GitHub Repository](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO11 Official Documentation](https://docs.ultralytics.com/models/yolo11/) +- **Docs:** [YOLO11 Official Documentation](https://docs.ultralytics.com/models/yolo11) [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } !!! tip "Did you know?" - Ultralytics YOLO11 supports more than just object detection. Out of the box, you can perform [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection using the exact same API. + Ultralytics YOLO11 supports more than just object detection. Out of the box, you can perform [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection using the exact same API. ## Architectural and Performance Comparison -When comparing these two detectors, we must look beyond the raw numbers and understand how their architectural choices impact real-world [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/). +When comparing these two detectors, we must look beyond the raw numbers and understand how their architectural choices impact real-world [model deployment](https://docs.ultralytics.com/guides/model-deployment-options). ### PP-YOLOE+ Architecture -PP-YOLOE+ relies heavily on the [PaddlePaddle framework](https://github.com/PaddlePaddle/Paddle). It introduces a powerful anchor-free paradigm, utilizing a RepResNet backbone and a modified Path Aggregation Network (PAN). The "+" variant improved upon its predecessor by incorporating large-scale dataset pre-training (like [Objects365](https://docs.ultralytics.com/datasets/detect/objects365/)) and an improved TaskAlignedAssigner. While it achieves high [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map), the hard dependency on PaddlePaddle can introduce friction for teams accustomed to PyTorch or TensorFlow environments. +PP-YOLOE+ relies heavily on the [PaddlePaddle framework](https://github.com/PaddlePaddle/Paddle). It introduces a powerful anchor-free paradigm, utilizing a RepResNet backbone and a modified Path Aggregation Network (PAN). The "+" variant improved upon its predecessor by incorporating large-scale dataset pre-training (like [Objects365](https://docs.ultralytics.com/datasets/detect/objects365)) and an improved TaskAlignedAssigner. While it achieves high [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map), the hard dependency on PaddlePaddle can introduce friction for teams accustomed to PyTorch or TensorFlow environments. ### YOLO11 Architecture -Ultralytics YOLO11 is built natively on [PyTorch](https://pytorch.org/), the industry standard for modern deep learning. Its architecture focuses heavily on a **Performance Balance**, achieving a favorable trade-off between speed and accuracy suitable for diverse real-world deployment scenarios. YOLO11 features an optimized C2f module for better gradient flow and a decoupled head that efficiently handles classification and regression tasks separately. Furthermore, YOLO11 is engineered for lower memory requirements, boasting significantly lower memory usage during training and inference compared to complex transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +Ultralytics YOLO11 is built natively on [PyTorch](https://pytorch.org/), the industry standard for modern deep learning. Its architecture focuses heavily on a **Performance Balance**, achieving a favorable trade-off between speed and accuracy suitable for diverse real-world deployment scenarios. YOLO11 features an optimized C2f module for better gradient flow and a decoupled head that efficiently handles classification and regression tasks separately. Furthermore, YOLO11 is engineered for lower memory requirements, boasting significantly lower memory usage during training and inference compared to complex transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). ### Performance Metrics Table @@ -94,30 +94,30 @@ PP-YOLOE+ is a strong choice for: YOLO11 is recommended for: -- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) where reliability and active maintenance are paramount. -- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [OBB](https://docs.ultralytics.com/tasks/obb/) within a single unified framework. -- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python/). +- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) where reliability and active maintenance are paramount. +- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [OBB](https://docs.ultralytics.com/tasks/obb) within a single unified framework. +- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python). ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage While academic benchmarks are important, the long-term success of an AI project relies heavily on the ecosystem surrounding the model. The [Ultralytics Platform](https://platform.ultralytics.com) offers distinct advantages for developers and enterprises alike. -1. **Ease of Use:** Ultralytics abstracts away the complexities of deep learning. The streamlined user experience and simple Python API allow developers to [train custom models](https://docs.ultralytics.com/modes/train/) with just a few lines of code. This contrasts with the complex configuration files often required by PP-YOLOE+. -2. **Well-Maintained Ecosystem:** Unlike many research-only repositories, the Ultralytics ecosystem is actively developed. It boasts strong community support, frequent updates, and extensive integration with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) and [Comet ML](https://docs.ultralytics.com/integrations/comet/). -3. **Versatility:** YOLO11 provides a single, unified framework for multiple [computer vision tasks](https://docs.ultralytics.com/tasks/), eliminating the need to learn different libraries for classification, segmentation, or bounding box detection. -4. **Training Efficiency:** The efficient training processes of YOLO models save both time and compute costs. By leveraging pre-trained weights on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/), models converge rapidly even on consumer-grade hardware. +1. **Ease of Use:** Ultralytics abstracts away the complexities of deep learning. The streamlined user experience and simple Python API allow developers to [train custom models](https://docs.ultralytics.com/modes/train) with just a few lines of code. This contrasts with the complex configuration files often required by PP-YOLOE+. +2. **Well-Maintained Ecosystem:** Unlike many research-only repositories, the Ultralytics ecosystem is actively developed. It boasts strong community support, frequent updates, and extensive integration with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) and [Comet ML](https://docs.ultralytics.com/integrations/comet). +3. **Versatility:** YOLO11 provides a single, unified framework for multiple [computer vision tasks](https://docs.ultralytics.com/tasks), eliminating the need to learn different libraries for classification, segmentation, or bounding box detection. +4. **Training Efficiency:** The efficient training processes of YOLO models save both time and compute costs. By leveraging pre-trained weights on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco), models converge rapidly even on consumer-grade hardware. ### Training Code Comparison -To illustrate the ease of use, here is how you train a state-of-the-art YOLO11 model. It handles all data [augmentation](https://docs.ultralytics.com/reference/data/augment/), logging, and hardware orchestration automatically: +To illustrate the ease of use, here is how you train a state-of-the-art YOLO11 model. It handles all data [augmentation](https://docs.ultralytics.com/reference/data/augment), logging, and hardware orchestration automatically: ```python from ultralytics import YOLO @@ -141,16 +141,16 @@ While YOLO11 remains an exceptionally powerful tool, the field of AI moves rapid YOLO26 introduces several groundbreaking innovations: -- **End-to-End NMS-Free Design:** Building on concepts first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 is natively end-to-end. It completely eliminates Non-Maximum Suppression (NMS) post-processing, making deployment vastly simpler and significantly reducing latency variability. -- **Up to 43% Faster CPU Inference:** By strategically removing Distribution Focal Loss (DFL), the model becomes much lighter. This optimization makes it the premier choice for [edge computing](https://docs.ultralytics.com/integrations/edge-tpu/) and low-power IoT devices. +- **End-to-End NMS-Free Design:** Building on concepts first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 is natively end-to-end. It completely eliminates Non-Maximum Suppression (NMS) post-processing, making deployment vastly simpler and significantly reducing latency variability. +- **Up to 43% Faster CPU Inference:** By strategically removing Distribution Focal Loss (DFL), the model becomes much lighter. This optimization makes it the premier choice for [edge computing](https://docs.ultralytics.com/integrations/edge-tpu) and low-power IoT devices. - **MuSGD Optimizer:** YOLO26 brings LLM training innovations to computer vision. Using the MuSGD optimizer (a hybrid of SGD and Muon), it achieves highly stable training dynamics and faster convergence. -- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, a critical feature for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and aerial surveillance. +- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, a critical feature for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and aerial surveillance. ## Conclusion and Real-World Applications When deciding between PP-YOLOE+ and YOLO11 (or the newer YOLO26), the choice hinges on your deployment ecosystem. -**PP-YOLOE+** shines in specific industrial environments, particularly in Asian manufacturing hubs where the hardware is deeply integrated with the Baidu technology stack and the [PaddlePaddle library](https://docs.ultralytics.com/integrations/paddlepaddle/). It is excellent for static image analysis where maximum mAP is the sole priority. +**PP-YOLOE+** shines in specific industrial environments, particularly in Asian manufacturing hubs where the hardware is deeply integrated with the Baidu technology stack and the [PaddlePaddle library](https://docs.ultralytics.com/integrations/paddlepaddle). It is excellent for static image analysis where maximum mAP is the sole priority. **YOLO11** and **YOLO26**, however, offer a much more versatile and developer-friendly approach. Their lower parameter count and high speeds make them ideal for: @@ -158,4 +158,4 @@ When deciding between PP-YOLOE+ and YOLO11 (or the newer YOLO26), the choice hin - **Autonomous Robotics:** Enabling [high-speed obstacle avoidance](https://www.ultralytics.com/blog/from-algorithms-to-automation-ais-role-in-robotics) on resource-constrained embedded devices. - **Security and Surveillance:** Providing robust, multi-task analysis (like tracking and pose estimation) in single, highly efficient inference passes. -For modern AI engineers looking for reliability, extensive community support, and straightforward deployment pipelines to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), the Ultralytics ecosystem remains the undisputed choice. +For modern AI engineers looking for reliability, extensive community support, and straightforward deployment pipelines to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), the Ultralytics ecosystem remains the undisputed choice. diff --git a/docs/en/compare/pp-yoloe-vs-yolo26.md b/docs/en/compare/pp-yoloe-vs-yolo26.md index 6f5e7cae611..18d095476d1 100644 --- a/docs/en/compare/pp-yoloe-vs-yolo26.md +++ b/docs/en/compare/pp-yoloe-vs-yolo26.md @@ -8,7 +8,7 @@ keywords: PP-YOLOE+, YOLO26, object detection, model comparison, computer vision The landscape of real-time computer vision has seen tremendous growth, driven by the need for scalable, efficient, and highly accurate object detection models. Two standout architectures in this space are **PP-YOLOE+**, a powerful detector from the [PaddlePaddle ecosystem](https://github.com/PaddlePaddle/PaddleDetection/), and **[Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26)**, the latest state-of-the-art model redefining edge deployment and training efficiency. -This comprehensive guide compares these two models, highlighting their architectures, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/), training methodologies, and ideal use cases to help you make an informed decision for your next AI project. +This comprehensive guide compares these two models, highlighting their architectures, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics), training methodologies, and ideal use cases to help you make an informed decision for your next AI project. @@ -36,7 +36,7 @@ Understanding the origins and design philosophies behind these models provides c - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** January 14, 2026 - **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26/) +- **Docs:** [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26) [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -52,15 +52,15 @@ Released in early 2026, **Ultralytics YOLO26** completely reimagines the real-ti Key YOLO26 innovations include: -- **End-to-End NMS-Free Design:** YOLO26 is natively end-to-end, completely eliminating the need for Non-Maximum Suppression ([NMS](https://www.ultralytics.com/glossary/non-maximum-suppression-nms)) post-processing. This breakthrough, first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), ensures consistent inference latency regardless of scene crowding, making deployment significantly simpler. +- **End-to-End NMS-Free Design:** YOLO26 is natively end-to-end, completely eliminating the need for Non-Maximum Suppression ([NMS](https://www.ultralytics.com/glossary/non-maximum-suppression-nms)) post-processing. This breakthrough, first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), ensures consistent inference latency regardless of scene crowding, making deployment significantly simpler. - **DFL Removal:** By removing Distribution Focal Loss (DFL), YOLO26 drastically simplifies its output head. This results in far better compatibility with edge devices and microcontrollers. - **Up to 43% Faster CPU Inference:** Thanks to the DFL removal and structural optimizations, YOLO26 is heavily optimized for environments without dedicated GPUs, achieving up to 43% faster inference speeds on CPUs compared to [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11). - **MuSGD Optimizer:** Inspired by advanced LLM training techniques like those from [Moonshot AI](https://www.moonshot.ai/), YOLO26 introduces a hybrid of SGD and Muon. This brings unparalleled training stability and faster convergence to computer vision tasks. -- **ProgLoss + STAL:** Advanced loss functions specifically target and improve small-object recognition, which is critical for [drone operations](https://docs.ultralytics.com/datasets/detect/visdrone/) and IoT edge sensors. +- **ProgLoss + STAL:** Advanced loss functions specifically target and improve small-object recognition, which is critical for [drone operations](https://docs.ultralytics.com/datasets/detect/visdrone) and IoT edge sensors. !!! tip "Task-Specific Improvements in YOLO26" - Beyond standard bounding boxes, YOLO26 introduces specific upgrades across all vision tasks. It uses semantic segmentation loss and multi-scale prototyping for [Segmentation](https://docs.ultralytics.com/tasks/segment/), Residual Log-Likelihood Estimation (RLE) for [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), and a specialized angle loss to resolve boundary issues in [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. + Beyond standard bounding boxes, YOLO26 introduces specific upgrades across all vision tasks. It uses semantic segmentation loss and multi-scale prototyping for [Segmentation](https://docs.ultralytics.com/tasks/segment), Residual Log-Likelihood Estimation (RLE) for [Pose Estimation](https://docs.ultralytics.com/tasks/pose), and a specialized angle loss to resolve boundary issues in [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. ## Performance and Metrics @@ -84,16 +84,16 @@ _Note: Bold values highlight the best-performing metrics across all models._ ### Analysis -- **Memory Requirements and Efficiency:** YOLO26 requires significantly fewer parameters and FLOPs to achieve higher mAP scores. For example, the YOLO26n (Nano) model achieves a 40.9 mAP with only 2.4M parameters, outperforming the PP-YOLOE+t model while being roughly half the size. This translates to lower memory usage during both [training](https://docs.ultralytics.com/modes/train/) and deployment. -- **Inference Speed:** When exported using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), YOLO26 dominates the latency metrics. The removal of NMS ensures that the 1.7ms inference time on a T4 GPU remains perfectly stable, whereas PP-YOLOE+ relies on potentially variable post-processing times. +- **Memory Requirements and Efficiency:** YOLO26 requires significantly fewer parameters and FLOPs to achieve higher mAP scores. For example, the YOLO26n (Nano) model achieves a 40.9 mAP with only 2.4M parameters, outperforming the PP-YOLOE+t model while being roughly half the size. This translates to lower memory usage during both [training](https://docs.ultralytics.com/modes/train) and deployment. +- **Inference Speed:** When exported using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), YOLO26 dominates the latency metrics. The removal of NMS ensures that the 1.7ms inference time on a T4 GPU remains perfectly stable, whereas PP-YOLOE+ relies on potentially variable post-processing times. ## The Ultralytics Advantage: Ecosystem and Ease of Use While raw metrics are important, the developer experience often dictates project success. The **[Ultralytics Platform](https://platform.ultralytics.com/)** provides a well-maintained ecosystem that completely outclasses older frameworks. 1. **Ease of Use:** Ultralytics abstracts away complex boilerplate code. Training YOLO26 takes only a few lines of Python, avoiding the dense configuration files required by PP-YOLOE+. -2. **Versatility:** PP-YOLOE+ is primarily an [object detection](https://docs.ultralytics.com/tasks/detect/) architecture. YOLO26 offers out-of-the-box support for segmentation, classification, pose estimation, and OBB. -3. **Training Efficiency:** Ultralytics YOLO models require vastly lower CUDA memory compared to bulky [transformer models](https://www.ultralytics.com/glossary/transformer) like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) or older architectures, enabling researchers to train state-of-the-art models on consumer-grade hardware. +2. **Versatility:** PP-YOLOE+ is primarily an [object detection](https://docs.ultralytics.com/tasks/detect) architecture. YOLO26 offers out-of-the-box support for segmentation, classification, pose estimation, and OBB. +3. **Training Efficiency:** Ultralytics YOLO models require vastly lower CUDA memory compared to bulky [transformer models](https://www.ultralytics.com/glossary/transformer) like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) or older architectures, enabling researchers to train state-of-the-art models on consumer-grade hardware. !!! note "Other Ultralytics Models" @@ -133,7 +133,7 @@ print(f"Model successfully exported to: {export_path}") ### When to Choose YOLO26 - **Edge Computing and IoT:** The **43% faster CPU inference** and DFL removal make YOLO26 the uncontested champion for deployment on Raspberry Pis, mobile phones, and embedded devices. -- **Crowded Scenes and Smart Cities:** The **End-to-End NMS-Free** architecture guarantees stable latency in dense environments like [parking management](https://docs.ultralytics.com/guides/parking-management/) and traffic monitoring, where traditional NMS would cause bottlenecks. +- **Crowded Scenes and Smart Cities:** The **End-to-End NMS-Free** architecture guarantees stable latency in dense environments like [parking management](https://docs.ultralytics.com/guides/parking-management) and traffic monitoring, where traditional NMS would cause bottlenecks. - **Multi-Task Projects:** If your pipeline requires tracking objects, estimating human poses, or generating pixel-perfect masks, YOLO26 handles it all within a single, unified Python package. ## Conclusion diff --git a/docs/en/compare/pp-yoloe-vs-yolov10.md b/docs/en/compare/pp-yoloe-vs-yolov10.md index ed5c305e30c..38676deb791 100644 --- a/docs/en/compare/pp-yoloe-vs-yolov10.md +++ b/docs/en/compare/pp-yoloe-vs-yolov10.md @@ -15,7 +15,7 @@ The landscape of computer vision is constantly evolving, with new models pushing ## Introduction to the Models -Choosing the right foundation for your [computer vision projects](https://docs.ultralytics.com/guides/steps-of-a-cv-project/) requires a deep understanding of each model's architectural trade-offs, deployment constraints, and ecosystem support. +Choosing the right foundation for your [computer vision projects](https://docs.ultralytics.com/guides/steps-of-a-cv-project) requires a deep understanding of each model's architectural trade-offs, deployment constraints, and ecosystem support. ### PP-YOLOE+ Overview @@ -28,7 +28,7 @@ Developed by the PaddlePaddle Authors at Baidu, PP-YOLOE+ is an evolutionary ste - **GitHub:** [PaddleDetection Repository](https://github.com/PaddlePaddle/PaddleDetection/) - **Docs:** [PP-YOLOE+ Official Documentation](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.8.1/configs/ppyoloe/README.md) -**Strengths:** PP-YOLOE+ excels in environments deeply integrated with the [PaddlePaddle framework](https://docs.ultralytics.com/integrations/paddlepaddle/). It introduces an advanced CSPRepResNet backbone and relies on a powerful label assignment strategy (TAL) to achieve impressive [mean Average Precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics/). It is highly optimized for deployment on server-grade GPUs common in industrial applications across Asia. +**Strengths:** PP-YOLOE+ excels in environments deeply integrated with the [PaddlePaddle framework](https://docs.ultralytics.com/integrations/paddlepaddle). It introduces an advanced CSPRepResNet backbone and relies on a powerful label assignment strategy (TAL) to achieve impressive [mean Average Precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics). It is highly optimized for deployment on server-grade GPUs common in industrial applications across Asia. **Weaknesses:** The primary drawback of PP-YOLOE+ is its heavy reliance on the PaddlePaddle ecosystem, which can be less intuitive for developers accustomed to PyTorch. Additionally, it requires traditional Non-Maximum Suppression (NMS) for post-processing, which adds latency and deployment complexity. @@ -43,17 +43,17 @@ Released by researchers at Tsinghua University, YOLOv10 brought a significant ar - **Date:** 2024-05-23 - **Arxiv:** [https://arxiv.org/abs/2405.14458](https://arxiv.org/abs/2405.14458) - **GitHub:** [YOLOv10 Repository](https://github.com/THU-MIG/yolov10) -- **Docs:** [YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10/) +- **Docs:** [YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10) -**Strengths:** The standout feature of YOLOv10 is its consistent dual assignments for NMS-free training. This means the model natively predicts bounding boxes without requiring a secondary filtering step, making [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/) much simpler and faster on [edge devices](https://www.ultralytics.com/glossary/edge-ai). It achieves an excellent balance of low parameter count and high accuracy. +**Strengths:** The standout feature of YOLOv10 is its consistent dual assignments for NMS-free training. This means the model natively predicts bounding boxes without requiring a secondary filtering step, making [model deployment](https://docs.ultralytics.com/guides/model-deployment-options) much simpler and faster on [edge devices](https://www.ultralytics.com/glossary/edge-ai). It achieves an excellent balance of low parameter count and high accuracy. -**Weaknesses:** While highly efficient for standard 2D [object detection](https://docs.ultralytics.com/tasks/detect/), YOLOv10 lacks native support for other vital computer vision tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [pose estimation](https://docs.ultralytics.com/tasks/pose/), limiting its versatility in complex, multi-task pipelines. +**Weaknesses:** While highly efficient for standard 2D [object detection](https://docs.ultralytics.com/tasks/detect), YOLOv10 lacks native support for other vital computer vision tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [pose estimation](https://docs.ultralytics.com/tasks/pose), limiting its versatility in complex, multi-task pipelines. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } !!! tip "Considering Advanced Alternatives?" - If you are exploring the latest innovations in real-time detection, consider reading our guide on [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) for high-accuracy vision applications. + If you are exploring the latest innovations in real-time detection, consider reading our guide on [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr) for high-accuracy vision applications. ## Performance and Metrics Comparison @@ -76,7 +76,7 @@ Understanding how these models perform under standardized benchmarks is crucial ### Technical Analysis -When analyzing the data, a few key trends emerge. The YOLOv10 nano and small models aggressively target edge efficiency, with YOLOv10n boasting a mere 2.3 million parameters and 6.7B FLOPs. This lightweight design, combined with its NMS-free architecture, drastically reduces latency on platforms utilizing [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino/). +When analyzing the data, a few key trends emerge. The YOLOv10 nano and small models aggressively target edge efficiency, with YOLOv10n boasting a mere 2.3 million parameters and 6.7B FLOPs. This lightweight design, combined with its NMS-free architecture, drastically reduces latency on platforms utilizing [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino). Conversely, PP-YOLOE+ demonstrates strong capability in the larger weight classes, with its X-large variant marginally edging out YOLOv10x in mAP (54.7% vs 54.4%). However, this comes at the cost of nearly double the parameter count (98.42M vs 56.9M), making YOLOv10x the significantly more efficient model for memory-constrained environments. @@ -84,7 +84,7 @@ Conversely, PP-YOLOE+ demonstrates strong capability in the larger weight classe While both PP-YOLOE+ and YOLOv10 offer compelling technical achievements, modern ML engineering demands more than just a raw architecture; it requires a [well-maintained ecosystem](https://www.ultralytics.com/about). -Ultralytics provides an industry-leading Python SDK that dramatically simplifies [data collection and annotation](https://docs.ultralytics.com/guides/data-collection-and-annotation/), training, and deployment. Compared to heavy research frameworks or older transformer models, Ultralytics architectures require a fraction of the CUDA memory during training, allowing for larger batch sizes and faster iterations. Furthermore, the Ultralytics suite offers immense versatility—supporting [image classification](https://docs.ultralytics.com/tasks/classify/), [OBB (Oriented Bounding Box)](https://docs.ultralytics.com/tasks/obb/), and robust object tracking right out of the box. +Ultralytics provides an industry-leading Python SDK that dramatically simplifies [data collection and annotation](https://docs.ultralytics.com/guides/data-collection-and-annotation), training, and deployment. Compared to heavy research frameworks or older transformer models, Ultralytics architectures require a fraction of the CUDA memory during training, allowing for larger batch sizes and faster iterations. Furthermore, the Ultralytics suite offers immense versatility—supporting [image classification](https://docs.ultralytics.com/tasks/classify), [OBB (Oriented Bounding Box)](https://docs.ultralytics.com/tasks/obb), and robust object tracking right out of the box. ### Enter YOLO26: The Next Generation @@ -144,11 +144,11 @@ YOLOv10 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Conclusion diff --git a/docs/en/compare/pp-yoloe-vs-yolov5.md b/docs/en/compare/pp-yoloe-vs-yolov5.md index 628da82ee67..4c8e1b4dd34 100644 --- a/docs/en/compare/pp-yoloe-vs-yolov5.md +++ b/docs/en/compare/pp-yoloe-vs-yolov5.md @@ -34,7 +34,7 @@ Both models stem from highly capable engineering teams but target slightly diffe - Organization: [Ultralytics](https://www.ultralytics.com/) - Date: 2020-06-26 - GitHub: [https://github.com/ultralytics/yolov5](https://github.com/ultralytics/yolov5) -- Docs: [https://docs.ultralytics.com/models/yolov5/](https://docs.ultralytics.com/models/yolov5/) +- Docs: [https://docs.ultralytics.com/models/yolov5/](https://docs.ultralytics.com/models/yolov5) [Learn more about YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5){ .md-button } @@ -74,7 +74,7 @@ Evaluating these models requires looking at the trade-off between mean Average P | YOLOv5l | 640 | 49.0 | 408.4 | 6.61 | 53.2 | 135.0 | | YOLOv5x | 640 | 50.7 | 763.2 | 11.89 | 97.2 | 246.4 | -While PP-YOLOE+ achieves highly competitive mAP scores at the larger scales (such as the X variant), **YOLOv5 provides superior speed and lower parameter counts** at the smaller end of the spectrum. The YOLOv5 Nano (`YOLOv5n`) requires a mere 2.6 million parameters, making it highly suitable for constrained edge devices where memory requirements are strict. Furthermore, training YOLO models typically consumes less CUDA memory compared to heavy transformer-based alternatives like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +While PP-YOLOE+ achieves highly competitive mAP scores at the larger scales (such as the X variant), **YOLOv5 provides superior speed and lower parameter counts** at the smaller end of the spectrum. The YOLOv5 Nano (`YOLOv5n`) requires a mere 2.6 million parameters, making it highly suitable for constrained edge devices where memory requirements are strict. Furthermore, training YOLO models typically consumes less CUDA memory compared to heavy transformer-based alternatives like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). ## The Ultralytics Advantage @@ -82,11 +82,11 @@ When choosing an architecture, raw metrics are only part of the equation. The de ### Unmatched Ease of Use -The [Python API](https://docs.ultralytics.com/usage/python/) for Ultralytics abstracts away complex boilerplate code. Developers can initiate training, validate performance, and deploy models seamlessly. The documentation is extensive, highly maintained, and supported by a massive global open-source community. +The [Python API](https://docs.ultralytics.com/usage/python) for Ultralytics abstracts away complex boilerplate code. Developers can initiate training, validate performance, and deploy models seamlessly. The documentation is extensive, highly maintained, and supported by a massive global open-source community. ### Versatility Across Tasks -While PP-YOLOE+ is a dedicated object detector, the Ultralytics ecosystem allows users to tackle multiple computer vision tasks under a single unified API. With YOLOv5, and its successors, you can effortlessly transition from standard bounding boxes to [Image Segmentation](https://docs.ultralytics.com/tasks/segment/) and classification workflows. +While PP-YOLOE+ is a dedicated object detector, the Ultralytics ecosystem allows users to tackle multiple computer vision tasks under a single unified API. With YOLOv5, and its successors, you can effortlessly transition from standard bounding boxes to [Image Segmentation](https://docs.ultralytics.com/tasks/segment) and classification workflows. ### Code Example: Training YOLOv5 @@ -112,7 +112,7 @@ predictions[0].show() If your organization is deeply embedded within the Baidu software stack or relies heavily on specialized hardware that mandates the PaddlePaddle framework, PP-YOLOE+ is a solid performer. It is frequently utilized in specialized manufacturing pipelines across Asia where legacy integration with Paddle exists. **When to choose YOLOv5:** -For the vast majority of international developers, researchers, and enterprises, YOLOv5 remains a powerhouse. Its PyTorch roots mean it is instantly compatible with tools like [Weights & Biases](https://wandb.ai/site) for tracking, and it exports cleanly to [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) for NVIDIA GPU acceleration or CoreML for Apple devices. It excels in diverse fields ranging from agricultural crop monitoring to high-speed drone navigation. +For the vast majority of international developers, researchers, and enterprises, YOLOv5 remains a powerhouse. Its PyTorch roots mean it is instantly compatible with tools like [Weights & Biases](https://wandb.ai/site) for tracking, and it exports cleanly to [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) for NVIDIA GPU acceleration or CoreML for Apple devices. It excels in diverse fields ranging from agricultural crop monitoring to high-speed drone navigation. ## The Future of Detection: Ultralytics YOLO26 @@ -125,6 +125,6 @@ While YOLOv5 is an iconic model, the frontier of computer vision has advanced. F - **End-to-End NMS-Free Design:** YOLO26 eliminates Non-Maximum Suppression post-processing entirely. This reduces latency variability and simplifies the deployment pipeline drastically. - **Up to 43% Faster CPU Inference:** By strategically removing Distribution Focal Loss (DFL), YOLO26 dramatically increases speed on edge devices without GPUs. - **MuSGD Optimizer:** Inspired by leading Large Language Models, this hybrid optimizer stabilizes training dynamics and allows for much faster convergence on custom datasets. -- **Task-Specific Enhancements:** Features advanced loss functions like ProgLoss and STAL, yielding unprecedented accuracy on tiny objects. It natively supports [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection for aerial imagery. +- **Task-Specific Enhancements:** Features advanced loss functions like ProgLoss and STAL, yielding unprecedented accuracy on tiny objects. It natively supports [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection for aerial imagery. -If you are exploring state-of-the-art vision models, you may also be interested in comparing the previous generation [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or transformer-based approaches like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). Ultimately, the robust ecosystem, combined with cutting-edge architectural advancements, cements Ultralytics as the premier choice for modern computer vision tasks. +If you are exploring state-of-the-art vision models, you may also be interested in comparing the previous generation [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or transformer-based approaches like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). Ultimately, the robust ecosystem, combined with cutting-edge architectural advancements, cements Ultralytics as the premier choice for modern computer vision tasks. diff --git a/docs/en/compare/pp-yoloe-vs-yolov6.md b/docs/en/compare/pp-yoloe-vs-yolov6.md index 2d5697c7997..ff52ebb396a 100644 --- a/docs/en/compare/pp-yoloe-vs-yolov6.md +++ b/docs/en/compare/pp-yoloe-vs-yolov6.md @@ -6,7 +6,7 @@ keywords: PP-YOLOE+, YOLOv6-3.0, object detection, model comparison, machine lea # Navigating Object Detection: PP-YOLOE+ vs YOLOv6-3.0 -The field of real-time [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) has expanded rapidly, leading to highly specialized architectures optimized for diverse deployment scenarios. Developers frequently compare [PP-YOLOE+](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.8.1/configs/ppyoloe/README.md) and [YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6/) when building applications that require a balance of high throughput and reliable accuracy. Both models brought substantial architectural improvements to the table upon their releases, focusing on enhancing inference speeds for industrial and edge applications. +The field of real-time [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) has expanded rapidly, leading to highly specialized architectures optimized for diverse deployment scenarios. Developers frequently compare [PP-YOLOE+](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.8.1/configs/ppyoloe/README.md) and [YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6) when building applications that require a balance of high throughput and reliable accuracy. Both models brought substantial architectural improvements to the table upon their releases, focusing on enhancing inference speeds for industrial and edge applications. Before diving into the detailed architectural breakdowns, explore the chart below to visualize how these models perform relative to one another in terms of speed and accuracy. @@ -29,7 +29,7 @@ Developed by the [PaddlePaddle Authors](https://github.com/PaddlePaddle/PaddleDe PP-YOLOE+ introduced several critical enhancements over the original PP-YOLOE design. It leverages a powerful CSPRepResNet backbone, which efficiently balances computational cost with feature extraction capabilities. Furthermore, it incorporates an advanced [feature pyramid network (FPN)](https://www.ultralytics.com/glossary/feature-pyramid-network-fpn) combined with a Path Aggregation Network (PAN) to ensure multi-scale feature fusion. One of its standout features is the ET-head (Efficient Task-aligned head), which significantly improves classification and localization coordination during [object detection](https://www.ultralytics.com/glossary/object-detection). -While PP-YOLOE+ achieves impressive [mean average precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map), its reliance on the PaddlePaddle ecosystem can sometimes present a steep learning curve for researchers accustomed to PyTorch-native workflows. This can slightly complicate the [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/) process when targeting heterogeneous edge devices that lack direct Paddle inference support. +While PP-YOLOE+ achieves impressive [mean average precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map), its reliance on the PaddlePaddle ecosystem can sometimes present a steep learning curve for researchers accustomed to PyTorch-native workflows. This can slightly complicate the [model deployment](https://docs.ultralytics.com/guides/model-deployment-options) process when targeting heterogeneous edge devices that lack direct Paddle inference support. !!! note "Deployment Context" @@ -49,11 +49,11 @@ Released by the Meituan Vision AI Department, YOLOv6-3.0 was explicitly engineer ### Architecture Highlights -YOLOv6-3.0 features an EfficientRep backbone specifically tailored to maximize hardware utilization, particularly on NVIDIA GPUs using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). The v3.0 update brought a Bi-directional Concatenation (BiC) module to the neck, enhancing spatial feature retention without severely bloating the parameter count. Additionally, it introduced an Anchor-Aided Training (AAT) strategy that fuses the benefits of anchor-based stability during [model training](https://docs.ultralytics.com/modes/train/) while maintaining a fast, anchor-free architecture during [real-time inference](https://www.ultralytics.com/glossary/real-time-inference). +YOLOv6-3.0 features an EfficientRep backbone specifically tailored to maximize hardware utilization, particularly on NVIDIA GPUs using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). The v3.0 update brought a Bi-directional Concatenation (BiC) module to the neck, enhancing spatial feature retention without severely bloating the parameter count. Additionally, it introduced an Anchor-Aided Training (AAT) strategy that fuses the benefits of anchor-based stability during [model training](https://docs.ultralytics.com/modes/train) while maintaining a fast, anchor-free architecture during [real-time inference](https://www.ultralytics.com/glossary/real-time-inference). However, because YOLOv6-3.0 is highly optimized for server-grade GPUs, its latency gains sometimes diminish when deployed on heavily constrained, CPU-only edge devices. This specialization means it excels in environments like offline video analytics but may trail behind dynamically optimized models on smaller, localized hardware. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Performance Comparison Table @@ -94,26 +94,26 @@ YOLOv6 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Advancing Beyond Legacy Models While PP-YOLOE+ and YOLOv6-3.0 offer targeted solutions, modern AI development requires versatile, memory-efficient workflows. This is where the [Ultralytics Platform](https://platform.ultralytics.com/) provides an unparalleled developer experience. With a unified Python API, you can seamlessly train, validate, and deploy cutting-edge models without the immense configuration overhead typically found in older research repositories. -Ultralytics models natively support a wide array of vision tasks beyond standard detection, including [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) extraction. Furthermore, they are highly optimized for lower memory usage during training—a stark contrast to [transformer-based models](https://www.ultralytics.com/glossary/transformer) like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) which generally demand massive GPU VRAM allocations. +Ultralytics models natively support a wide array of vision tasks beyond standard detection, including [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) extraction. Furthermore, they are highly optimized for lower memory usage during training—a stark contrast to [transformer-based models](https://www.ultralytics.com/glossary/transformer) like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) which generally demand massive GPU VRAM allocations. ### Discover YOLO26: The New Standard For organizations looking to deploy the ultimate state-of-the-art vision models, [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) (released in January 2026) redefines performance boundaries. It significantly outperforms older generations with several critical innovations: -- **End-to-End NMS-Free Design:** Building on concepts from [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 completely eliminates [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing. This natively end-to-end approach guarantees predictable, ultra-low latency inference, crucial for real-time safety systems. +- **End-to-End NMS-Free Design:** Building on concepts from [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 completely eliminates [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing. This natively end-to-end approach guarantees predictable, ultra-low latency inference, crucial for real-time safety systems. - **Up to 43% Faster CPU Inference:** Through the removal of Distribution Focal Loss (DFL) from the architecture, YOLO26 is radically optimized for edge computing and environments lacking dedicated GPU acceleration. -- **MuSGD Optimizer:** Integrating LLM training stability into vision models, this hybrid optimizer (inspired by Moonshot AI) enables rapid convergence and highly stable [custom training](https://docs.ultralytics.com/guides/custom-trainer/) sessions. -- **ProgLoss + STAL:** These advanced loss formulations deliver remarkable improvements in small-object recognition, vital for applications like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and crowded scene analysis. +- **MuSGD Optimizer:** Integrating LLM training stability into vision models, this hybrid optimizer (inspired by Moonshot AI) enables rapid convergence and highly stable [custom training](https://docs.ultralytics.com/guides/custom-trainer) sessions. +- **ProgLoss + STAL:** These advanced loss formulations deliver remarkable improvements in small-object recognition, vital for applications like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and crowded scene analysis. !!! tip "Future-Proof Your Pipelines" @@ -121,7 +121,7 @@ For organizations looking to deploy the ultimate state-of-the-art vision models, ### Seamless Implementation -Training and exporting state-of-the-art models using the [Ultralytics Python package](https://docs.ultralytics.com/usage/python/) is remarkably simple. The following example demonstrates how to train the latest YOLO26 model and export it to ONNX for rapid edge deployment: +Training and exporting state-of-the-art models using the [Ultralytics Python package](https://docs.ultralytics.com/usage/python) is remarkably simple. The following example demonstrates how to train the latest YOLO26 model and export it to ONNX for rapid edge deployment: ```python from ultralytics import YOLO diff --git a/docs/en/compare/pp-yoloe-vs-yolov7.md b/docs/en/compare/pp-yoloe-vs-yolov7.md index 3d4c00df9a9..b1cc8f99a88 100644 --- a/docs/en/compare/pp-yoloe-vs-yolov7.md +++ b/docs/en/compare/pp-yoloe-vs-yolov7.md @@ -39,9 +39,9 @@ Developed by Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao, YOLOv7 - **Date:** 2022-07-06 - **Arxiv:** [2207.02696](https://arxiv.org/abs/2207.02696) - **GitHub:** [YOLOv7 Repository](https://github.com/WongKinYiu/yolov7) -- **Docs:** [Ultralytics YOLOv7 Docs](https://docs.ultralytics.com/models/yolov7/) +- **Docs:** [Ultralytics YOLOv7 Docs](https://docs.ultralytics.com/models/yolov7) -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ## Architectural Innovations @@ -51,11 +51,11 @@ PP-YOLOE+ relies heavily on an anchor-free paradigm, making the deployment proce ### YOLOv7 Architecture -YOLOv7 took a different approach by introducing the Extended Efficient Layer Aggregation Network (E-ELAN). This architecture allows the network to learn more diverse features without destroying the original gradient path, leading to better convergence. YOLOv7 also heavily utilizes model re-parameterization—specifically, planned re-parameterized convolutions—which merges convolutional layers during inference to speed up execution without sacrificing accuracy. This makes YOLOv7 exceptionally strong in tasks like [multi-object tracking](https://www.ultralytics.com/glossary/multi-object-tracking-mot) and complex [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/). +YOLOv7 took a different approach by introducing the Extended Efficient Layer Aggregation Network (E-ELAN). This architecture allows the network to learn more diverse features without destroying the original gradient path, leading to better convergence. YOLOv7 also heavily utilizes model re-parameterization—specifically, planned re-parameterized convolutions—which merges convolutional layers during inference to speed up execution without sacrificing accuracy. This makes YOLOv7 exceptionally strong in tasks like [multi-object tracking](https://www.ultralytics.com/glossary/multi-object-tracking-mot) and complex [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system). !!! note "Ecosystem Differences" - While PP-YOLOE+ is tightly integrated with Baidu's PaddlePaddle framework, YOLOv7 was built in [PyTorch](https://www.ultralytics.com/glossary/pytorch), which historically offers a larger community and broader out-of-the-box compatibility with deployment pipelines like [ONNX](https://docs.ultralytics.com/integrations/onnx/) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). + While PP-YOLOE+ is tightly integrated with Baidu's PaddlePaddle framework, YOLOv7 was built in [PyTorch](https://www.ultralytics.com/glossary/pytorch), which historically offers a larger community and broader out-of-the-box compatibility with deployment pipelines like [ONNX](https://docs.ultralytics.com/integrations/onnx) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). ## Performance Analysis @@ -80,7 +80,7 @@ When training and deploying these models, the framework you choose is just as im - **Well-Maintained Ecosystem:** Ultralytics YOLO models benefit from a continually updated ecosystem, robust documentation, and an active community. - **Memory Requirements:** Ultralytics heavily optimizes data loading and training regimes. Training Ultralytics YOLO models typically requires far less CUDA memory compared to heavy transformer-based architectures, allowing developers to utilize larger [batch sizes](https://www.ultralytics.com/glossary/batch-size) on consumer-grade hardware. -- **Training Efficiency:** Leveraging robust [data augmentation strategies](https://docs.ultralytics.com/guides/yolo-data-augmentation/) and built-in hyperparameter tuning, Ultralytics ensures that models converge quickly with readily available pre-trained weights. +- **Training Efficiency:** Leveraging robust [data augmentation strategies](https://docs.ultralytics.com/guides/yolo-data-augmentation) and built-in hyperparameter tuning, Ultralytics ensures that models converge quickly with readily available pre-trained weights. ### Simple API Implementation @@ -105,7 +105,7 @@ While PP-YOLOE+ and YOLOv7 are milestones in object detection, the landscape of **Why YOLO26 Outperforms Older Architectures:** -- **End-to-End NMS-Free Design:** YOLO26 is natively end-to-end. By eliminating Non-Maximum Suppression (NMS) post-processing, it guarantees predictable, deterministic inference latency—a breakthrough first seen in [YOLOv10](https://docs.ultralytics.com/models/yolov10/). +- **End-to-End NMS-Free Design:** YOLO26 is natively end-to-end. By eliminating Non-Maximum Suppression (NMS) post-processing, it guarantees predictable, deterministic inference latency—a breakthrough first seen in [YOLOv10](https://docs.ultralytics.com/models/yolov10). - **DFL Removal:** The removal of Distribution Focal Loss simplifies the export process and significantly improves compatibility for low-power edge devices. - **Up to 43% Faster CPU Inference:** For scenarios lacking dedicated GPUs—such as [smart city IoT sensors](https://www.ultralytics.com/blog/computer-vision-ai-in-smart-cities)—YOLO26 is heavily optimized to run efficiently directly on CPUs. - **MuSGD Optimizer:** Inspired by advanced LLM training techniques (like Moonshot AI's Kimi K2), YOLO26 uses a hybrid of SGD and Muon for incredibly stable training and fast convergence. @@ -121,11 +121,11 @@ PP-YOLOE+ shines when you are deeply entrenched in the Baidu and PaddlePaddle ec ### When to use YOLOv7 -YOLOv7 remains an excellent choice for generic high-performance inference, particularly when deploying on NVIDIA hardware utilizing [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). Its integration into the PyTorch ecosystem makes it highly versatile for academic research and custom commercial pipelines, such as [real-time crowd management](https://www.ultralytics.com/blog/vision-ai-in-crowd-management) or complex [pose estimation](https://docs.ultralytics.com/tasks/pose/) tasks where structural integrity of the network is paramount. +YOLOv7 remains an excellent choice for generic high-performance inference, particularly when deploying on NVIDIA hardware utilizing [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). Its integration into the PyTorch ecosystem makes it highly versatile for academic research and custom commercial pipelines, such as [real-time crowd management](https://www.ultralytics.com/blog/vision-ai-in-crowd-management) or complex [pose estimation](https://docs.ultralytics.com/tasks/pose) tasks where structural integrity of the network is paramount. ### Other Models to Consider -Depending on your exact needs, you might also be interested in comparing these architectures against [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) for broad, production-ready flexibility, or [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) if your project requires the specific advantages of vision transformers over traditional convolutional networks. +Depending on your exact needs, you might also be interested in comparing these architectures against [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) for broad, production-ready flexibility, or [RT-DETR](https://docs.ultralytics.com/models/rtdetr) if your project requires the specific advantages of vision transformers over traditional convolutional networks. ## Conclusion diff --git a/docs/en/compare/pp-yoloe-vs-yolov8.md b/docs/en/compare/pp-yoloe-vs-yolov8.md index 8f4ce8abc42..c37a9a2a866 100644 --- a/docs/en/compare/pp-yoloe-vs-yolov8.md +++ b/docs/en/compare/pp-yoloe-vs-yolov8.md @@ -34,21 +34,21 @@ PP-YOLOE+ utilizes a CSPRepResNet backbone and an Efficient Task-aligned head (E ### Ultralytics YOLOv8 -Released as a massive leap forward by Ultralytics, YOLOv8 established a new state-of-the-art for [object detection](https://docs.ultralytics.com/tasks/detect/), bringing unparalleled ease of use, extreme versatility, and high-speed execution to the broader PyTorch developer community. +Released as a massive leap forward by Ultralytics, YOLOv8 established a new state-of-the-art for [object detection](https://docs.ultralytics.com/tasks/detect), bringing unparalleled ease of use, extreme versatility, and high-speed execution to the broader PyTorch developer community. - **Authors:** Glenn Jocher, Ayush Chaurasia, and Jing Qiu - **Organization:** [Ultralytics](https://www.ultralytics.com) - **Date:** 2023-01-10 - **GitHub:** [ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8/) +- **Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8) -YOLOv8 introduced a highly optimized, anchor-free detection head and a revamped C2f building block replacing the older C3 module. This design provides superior gradient flow and allows for incredibly fast [model training](https://docs.ultralytics.com/modes/train/). Beyond simple detection, YOLOv8 is a multi-task powerhouse, seamlessly supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) through the exact same user-friendly API. +YOLOv8 introduced a highly optimized, anchor-free detection head and a revamped C2f building block replacing the older C3 module. This design provides superior gradient flow and allows for incredibly fast [model training](https://docs.ultralytics.com/modes/train). Beyond simple detection, YOLOv8 is a multi-task powerhouse, seamlessly supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) through the exact same user-friendly API. [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } ## Performance and Metrics Comparison -A direct comparison of these architectures reveals varying trade-offs between sheer parameter size and inference latency. Below is the performance breakdown using the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). +A direct comparison of these architectures reveals varying trade-offs between sheer parameter size and inference latency. Below is the performance breakdown using the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -70,7 +70,7 @@ While the largest PP-YOLOE+x model slightly edges out YOLOv8x in mAP, it comes a When evaluating models, the surrounding ecosystem is as crucial as the raw architecture. PP-YOLOE+ demands navigating complex configuration files and dependencies specific to the PaddlePaddle framework. -Conversely, the Ultralytics experience is designed for maximum developer velocity. The well-maintained ecosystem boasts a simple [Python API](https://docs.ultralytics.com/usage/python/) and an incredibly active community. Furthermore, the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolov8) simplifies the entire ML pipeline, offering seamless dataset management, cloud training, and simple exports to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). +Conversely, the Ultralytics experience is designed for maximum developer velocity. The well-maintained ecosystem boasts a simple [Python API](https://docs.ultralytics.com/usage/python) and an incredibly active community. Furthermore, the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolov8) simplifies the entire ML pipeline, offering seamless dataset management, cloud training, and simple exports to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). !!! tip "Streamlined PyTorch Deployment" @@ -112,17 +112,17 @@ PP-YOLOE+ is a strong choice for: YOLOv8 is recommended for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Moving Beyond YOLOv8: The Dawn of YOLO26 @@ -130,12 +130,12 @@ While YOLOv8 remains a robust and reliable choice, developers looking for the ab YOLO26 brings several groundbreaking innovations that surpass both PP-YOLOE+ and previous YOLO generations (including [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11)): -- **End-to-End NMS-Free Design:** Building on concepts from [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 operates natively end-to-end. By eliminating [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing, it delivers consistent, ultra-low latency inference, regardless of how crowded the visual scene is. +- **End-to-End NMS-Free Design:** Building on concepts from [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 operates natively end-to-end. By eliminating [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing, it delivers consistent, ultra-low latency inference, regardless of how crowded the visual scene is. - **Up to 43% Faster CPU Inference:** Through the strategic removal of Distribution Focal Loss (DFL), YOLO26 significantly cuts down on processing overhead, making it drastically faster on edge CPUs—ideal for [smart city](https://www.ultralytics.com/blog/computer-vision-ai-in-smart-cities) and IoT applications where expensive GPUs aren't available. - **MuSGD Optimizer:** YOLO26 borrows innovations from Large Language Model (LLM) training. Its hybrid MuSGD optimizer brings unprecedented stability and faster convergence during training. -- **ProgLoss + STAL:** These advanced loss formulations vastly improve the detection of small and distant objects. This is a game-changer for drone operators monitoring [agricultural fields](https://docs.ultralytics.com/datasets/detect/visdrone/) or defect detection on fast-moving manufacturing lines. +- **ProgLoss + STAL:** These advanced loss formulations vastly improve the detection of small and distant objects. This is a game-changer for drone operators monitoring [agricultural fields](https://docs.ultralytics.com/datasets/detect/visdrone) or defect detection on fast-moving manufacturing lines. -For developers starting new computer vision initiatives, [YOLO26](https://docs.ultralytics.com/models/yolo26/) is the definitive recommendation. +For developers starting new computer vision initiatives, [YOLO26](https://docs.ultralytics.com/models/yolo26) is the definitive recommendation. ## Real-World Applications @@ -148,8 +148,8 @@ Choosing between these models often depends on your specific deployment reality: **Where Ultralytics YOLOv8 (and YOLO26) Excels:** -- **Dynamic Edge Computing:** From [NVIDIA Jetson devices](https://docs.ultralytics.com/guides/nvidia-jetson/) to basic Raspberry Pis, Ultralytics models provide the optimal balance of speed and lightweight memory footprints. -- **Multi-Task Pipelines:** If your application needs to evolve from simple bounding boxes to [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) for aerial imagery, or pose estimation for behavioral analysis, Ultralytics supports all tasks out-of-the-box. -- **Rapid Prototyping to Production:** The Ultralytics ecosystem empowers teams to iterate quickly. With pre-trained weights readily available, custom models can be spun up, trained, and deployed via the [Ultralytics Platform](https://docs.ultralytics.com/platform/) in a fraction of the time required by competing architectures. +- **Dynamic Edge Computing:** From [NVIDIA Jetson devices](https://docs.ultralytics.com/guides/nvidia-jetson) to basic Raspberry Pis, Ultralytics models provide the optimal balance of speed and lightweight memory footprints. +- **Multi-Task Pipelines:** If your application needs to evolve from simple bounding boxes to [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb) for aerial imagery, or pose estimation for behavioral analysis, Ultralytics supports all tasks out-of-the-box. +- **Rapid Prototyping to Production:** The Ultralytics ecosystem empowers teams to iterate quickly. With pre-trained weights readily available, custom models can be spun up, trained, and deployed via the [Ultralytics Platform](https://docs.ultralytics.com/platform) in a fraction of the time required by competing architectures. While PP-YOLOE+ offers competitive benchmarks, the unparalleled versatility, ease of use, and continual innovation—evidenced by the release of YOLO26—solidify Ultralytics models as the superior choice for modern developers and researchers alike. diff --git a/docs/en/compare/pp-yoloe-vs-yolov9.md b/docs/en/compare/pp-yoloe-vs-yolov9.md index 8b6dfba08ef..de899b6278a 100644 --- a/docs/en/compare/pp-yoloe-vs-yolov9.md +++ b/docs/en/compare/pp-yoloe-vs-yolov9.md @@ -6,9 +6,9 @@ keywords: PP-YOLOE+, YOLOv9, object detection, model comparison, computer vision # PP-YOLOE+ vs. YOLOv9: A Technical Deep Dive into Modern Object Detection -The landscape of real-time computer vision is constantly shifting, with researchers and developers continuously pushing the boundaries of accuracy and inference speed. When comparing [PP-YOLOE+](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.8.1/configs/ppyoloe/README.md) and [YOLOv9](https://docs.ultralytics.com/models/yolov9/), we are looking at two distinct philosophies in model architecture and ecosystem design. +The landscape of real-time computer vision is constantly shifting, with researchers and developers continuously pushing the boundaries of accuracy and inference speed. When comparing [PP-YOLOE+](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.8.1/configs/ppyoloe/README.md) and [YOLOv9](https://docs.ultralytics.com/models/yolov9), we are looking at two distinct philosophies in model architecture and ecosystem design. -This comprehensive technical comparison analyzes their architectural innovations, performance metrics, training methodologies, and ideal use cases to help you choose the right [object detection](https://docs.ultralytics.com/tasks/detect/) model for your next deployment. +This comprehensive technical comparison analyzes their architectural innovations, performance metrics, training methodologies, and ideal use cases to help you choose the right [object detection](https://docs.ultralytics.com/tasks/detect) model for your next deployment. @@ -17,7 +17,7 @@ This comprehensive technical comparison analyzes their architectural innovations ## Model Lineage and Technical Foundations -Understanding the origins and architectural choices of these models is crucial for determining their fit within your [computer vision projects](https://docs.ultralytics.com/guides/steps-of-a-cv-project/). +Understanding the origins and architectural choices of these models is crucial for determining their fit within your [computer vision projects](https://docs.ultralytics.com/guides/steps-of-a-cv-project). ### PP-YOLOE+ Overview @@ -45,15 +45,15 @@ Introduced by Chien-Yao Wang and Hong-Yuan Mark Liao from the Institute of Infor YOLOv9's major breakthrough is Programmable Gradient Information (PGI), which prevents data loss as features pass through deep neural networks. Combined with the Generalized Efficient Layer Aggregation Network (GELAN), YOLOv9 maximizes parameter efficiency and computational flow. Furthermore, it is natively integrated into the [Ultralytics ecosystem](https://docs.ultralytics.com/), making it highly accessible for both research and commercial applications. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } !!! note "Other Ultralytics Models" - If you are exploring state-of-the-art options, you might also be interested in [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [RT-DETR](https://docs.ultralytics.com/models/rtdetr/), which offer varying balances of transformer-based precision and real-time edge performance. + If you are exploring state-of-the-art options, you might also be interested in [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [RT-DETR](https://docs.ultralytics.com/models/rtdetr), which offer varying balances of transformer-based precision and real-time edge performance. ## Performance and Metrics Comparison -When analyzing raw performance, YOLOv9 demonstrates exceptional parameter efficiency. It achieves comparable or superior accuracy while requiring fewer parameters and FLOPs, translating to lower VRAM requirements during [model training](https://docs.ultralytics.com/modes/train/). +When analyzing raw performance, YOLOv9 demonstrates exceptional parameter efficiency. It achieves comparable or superior accuracy while requiring fewer parameters and FLOPs, translating to lower VRAM requirements during [model training](https://docs.ultralytics.com/modes/train). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -75,7 +75,7 @@ As seen in the table, YOLOv9c achieves a strong 53.0 mAP with significantly fewe The defining advantage of YOLOv9 lies in its seamless integration with the well-maintained Ultralytics ecosystem. While PP-YOLOE+ requires navigating complex PaddlePaddle configuration files, YOLOv9 benefits from a streamlined Python API. -The [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) allows developers to load pre-trained weights, manage [data augmentation](https://docs.ultralytics.com/reference/data/augment/), and initiate training with minimal boilerplate code. +The [Ultralytics Python API](https://docs.ultralytics.com/usage/python) allows developers to load pre-trained weights, manage [data augmentation](https://docs.ultralytics.com/reference/data/augment), and initiate training with minimal boilerplate code. ```python from ultralytics import YOLO @@ -93,11 +93,11 @@ results = model("https://ultralytics.com/images/bus.jpg") model.export(format="onnx") ``` -Furthermore, the Ultralytics ecosystem provides unmatched versatility. Beyond bounding box detection, the framework natively supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. This makes adapting your model to complex real-world pipelines incredibly efficient. +Furthermore, the Ultralytics ecosystem provides unmatched versatility. Beyond bounding box detection, the framework natively supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. This makes adapting your model to complex real-world pipelines incredibly efficient. !!! tip "Export Options" - Models trained using the Ultralytics framework can be exported to multiple formats, including [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino/), ensuring highly optimized inference across diverse hardware. + Models trained using the Ultralytics framework can be exported to multiple formats, including [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino), ensuring highly optimized inference across diverse hardware. ## Use Cases and Recommendations @@ -121,11 +121,11 @@ YOLOv9 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Looking Forward: The YOLO26 Advantage @@ -144,7 +144,7 @@ You can easily train and deploy YOLO26 models through the [Ultralytics Platform] Choosing between these architectures often comes down to your target deployment environment. -**PP-YOLOE+** is frequently deployed in industrial manufacturing centers, particularly in regions where the [PaddlePaddle integration](https://docs.ultralytics.com/integrations/paddlepaddle/) and Baidu's hardware stack are deeply embedded into enterprise infrastructure. It excels in static image analysis where absolute precision is prioritized over strict real-time constraints. +**PP-YOLOE+** is frequently deployed in industrial manufacturing centers, particularly in regions where the [PaddlePaddle integration](https://docs.ultralytics.com/integrations/paddlepaddle) and Baidu's hardware stack are deeply embedded into enterprise infrastructure. It excels in static image analysis where absolute precision is prioritized over strict real-time constraints. **YOLOv9** excels in dynamic environments requiring rapid [real-time inference](https://www.ultralytics.com/blog/real-time-inferences-in-vision-ai-solutions-are-making-an-impact). Its superior parameter efficiency makes it ideal for autonomous drone navigation and edge-based security systems. Furthermore, its lower VRAM consumption lowers the barrier to entry for researchers training on consumer-grade GPUs. diff --git a/docs/en/compare/pp-yoloe-vs-yolox.md b/docs/en/compare/pp-yoloe-vs-yolox.md index 42c2d33c743..5d36e408ae5 100644 --- a/docs/en/compare/pp-yoloe-vs-yolox.md +++ b/docs/en/compare/pp-yoloe-vs-yolox.md @@ -15,7 +15,7 @@ The landscape of [computer vision](https://www.ultralytics.com/glossary/computer ## Model Lineage and Details -Before diving into the technical architectures, it is helpful to contextualize the origins of both models. Each was developed to address specific bottlenecks in [object detection](https://docs.ultralytics.com/tasks/detect/), heavily influenced by their backing organizations. +Before diving into the technical architectures, it is helpful to contextualize the origins of both models. Each was developed to address specific bottlenecks in [object detection](https://docs.ultralytics.com/tasks/detect), heavily influenced by their backing organizations. **PP-YOLOE+ Details:** @@ -45,7 +45,7 @@ The core differences between these two detectors lie in their approach to featur **YOLOX** made waves in 2021 by successfully adapting the YOLO family to an **anchor-free** design. By removing anchor boxes, YOLOX significantly reduced the number of design parameters and heuristic tuning required for custom datasets. Furthermore, it introduced a decoupled head, which separates classification and localization tasks into distinct neural pathways. This separation resolved the inherent conflict between classifying an object and regressing its spatial coordinates, leading to faster convergence during training. -**PP-YOLOE+**, developed by Baidu, is heavily optimized for the [PaddlePaddle](https://docs.ultralytics.com/integrations/paddlepaddle/) ecosystem. It builds upon its predecessor, PP-YOLOv2, by introducing a dynamic label assignment strategy (TAL) and a novel backbone called CSPRepResNet. This backbone leverages structural re-parameterization, allowing the model to benefit from complex multi-branch architectures during training while seamlessly folding into a fast, single-path network for inference. +**PP-YOLOE+**, developed by Baidu, is heavily optimized for the [PaddlePaddle](https://docs.ultralytics.com/integrations/paddlepaddle) ecosystem. It builds upon its predecessor, PP-YOLOv2, by introducing a dynamic label assignment strategy (TAL) and a novel backbone called CSPRepResNet. This backbone leverages structural re-parameterization, allowing the model to benefit from complex multi-branch architectures during training while seamlessly folding into a fast, single-path network for inference. !!! tip "Structural Re-parameterization" @@ -80,7 +80,7 @@ Choosing between these frameworks often comes down to integration requirements a ### Where YOLOX Excels -Due to its anchor-free nature and availability of extreme edge variants, YOLOX is popular in [robotics](https://www.ultralytics.com/solutions/ai-in-robotics) and microcontroller deployment. Its simple post-processing pipeline allows for easier porting to customized NPU hardware formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) and [NCNN](https://docs.ultralytics.com/integrations/ncnn/). +Due to its anchor-free nature and availability of extreme edge variants, YOLOX is popular in [robotics](https://www.ultralytics.com/solutions/ai-in-robotics) and microcontroller deployment. Its simple post-processing pipeline allows for easier porting to customized NPU hardware formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) and [NCNN](https://docs.ultralytics.com/integrations/ncnn). ### Where PP-YOLOE+ Excels @@ -108,11 +108,11 @@ YOLOX is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Enter YOLO26 @@ -126,7 +126,7 @@ For teams looking to transition from isolated research repositories to productio - **Up to 43% Faster CPU Inference:** By strategically removing Distribution Focal Loss (DFL), YOLO26 achieves unparalleled inference speeds on CPU hardware, making it far superior for [edge computing](https://www.ultralytics.com/glossary/edge-computing) and low-power devices. - **MuSGD Optimizer:** Inspired by Moonshot AI’s Kimi K2, this hybrid optimizer brings LLM training stability to computer vision, ensuring much faster convergence and minimizing the memory requirements during training phases. - **ProgLoss + STAL:** These advanced loss functions deliver notable improvements in small-object recognition, a critical feature for [drone operations](https://www.ultralytics.com/blog/build-ai-powered-drone-applications-with-ultralytics-yolo11) and highly detailed aerial imagery. -- **Versatility:** While PP-YOLOE+ and YOLOX focus purely on detection, YOLO26 seamlessly handles [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) using the exact same intuitive syntax. +- **Versatility:** While PP-YOLOE+ and YOLOX focus purely on detection, YOLO26 seamlessly handles [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb) using the exact same intuitive syntax. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -158,4 +158,4 @@ model.export(format="engine") Both PP-YOLOE+ and YOLOX have earned their places in computer vision history, offering high accuracy and lightweight anchor-free designs, respectively. However, for organizations building the future of [AI in agriculture](https://www.ultralytics.com/solutions/ai-in-agriculture), smart cities, and retail, the continuous maintenance, ease of use, and native NMS-free architecture of **Ultralytics YOLO26** make it the undisputed choice. -If you are exploring alternative architectures for specific benchmarks, you may also find value in comparing the older [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or transformer-based options like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) via the comprehensive Ultralytics documentation. By migrating to the unified Ultralytics ecosystem, developers save invaluable time and resources while achieving state-of-the-art results on any edge or cloud deployment. +If you are exploring alternative architectures for specific benchmarks, you may also find value in comparing the older [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or transformer-based options like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) via the comprehensive Ultralytics documentation. By migrating to the unified Ultralytics ecosystem, developers save invaluable time and resources while achieving state-of-the-art results on any edge or cloud deployment. diff --git a/docs/en/compare/rtdetr-vs-damo-yolo.md b/docs/en/compare/rtdetr-vs-damo-yolo.md index b4c57ba8e15..adbc1aa3193 100644 --- a/docs/en/compare/rtdetr-vs-damo-yolo.md +++ b/docs/en/compare/rtdetr-vs-damo-yolo.md @@ -23,7 +23,7 @@ Understanding the core mechanics of these models is crucial for [machine learnin Building on the success of the original RT-DETR, RTDETRv2 utilizes a hybrid encoder and a [transformer decoder](https://arxiv.org/abs/1706.03762). This design allows the model to process global context highly effectively, making it exceptionally good at distinguishing between overlapping objects in dense scenes. The most significant advantage of this architecture is its native NMS-free (Non-Maximum Suppression) design. By eliminating the NMS post-processing step, RTDETRv2 streamlines the inference pipeline and ensures more stable latency across varying hardware configurations. -[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr){ .md-button } ### DAMO-YOLO: Advancing CNN Efficiency @@ -37,7 +37,7 @@ DAMO-YOLO, on the other hand, remains rooted in the highly successful CNN-based ## Performance and Metrics Comparison -When evaluating models for deployment, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/) such as mean Average Precision (mAP), inference speed, and parameter count are paramount. Below is a detailed comparison of the two model families. +When evaluating models for deployment, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics) such as mean Average Precision (mAP), inference speed, and parameter count are paramount. Below is a detailed comparison of the two model families. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -53,7 +53,7 @@ When evaluating models for deployment, [performance metrics](https://docs.ultral ### Analysis of Results -As seen in the table, the **RTDETRv2-x** achieves the highest accuracy with an mAPval of 54.3, showcasing the power of the transformer architecture on complex validations like the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). However, this comes at the cost of significantly higher parameters (76M) and FLOPs. +As seen in the table, the **RTDETRv2-x** achieves the highest accuracy with an mAPval of 54.3, showcasing the power of the transformer architecture on complex validations like the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). However, this comes at the cost of significantly higher parameters (76M) and FLOPs. Conversely, **DAMO-YOLOt** (Tiny) is exceptionally lightweight, requiring only 8.5M parameters, making it an incredibly fast option for environments where CUDA memory is severely restricted. DAMO-YOLO generally provides a favorable trade-off between speed and accuracy for legacy edge devices. @@ -86,13 +86,13 @@ results_yolo[0].show() !!! tip "Exporting Models for Production" - Using the Ultralytics API, you can seamlessly [export your trained models](https://docs.ultralytics.com/modes/export/) to formats like TensorRT, ONNX, or CoreML with a simple `model.export(format="engine")` command, drastically reducing deployment friction. + Using the Ultralytics API, you can seamlessly [export your trained models](https://docs.ultralytics.com/modes/export) to formats like TensorRT, ONNX, or CoreML with a simple `model.export(format="engine")` command, drastically reducing deployment friction. ## Ideal Use Cases Choosing between these architectures depends entirely on your specific project requirements: -- **RTDETRv2** excels in server-side processing where VRAM is abundant. Its global context awareness is perfect for [medical imaging](https://www.nature.com/subjects/medical-imaging?error=cookies_not_supported&code=fe3dedd7-5a9d-4206-9baf-f97b0ad990aa) and dense crowd analysis where occlusions are frequent. +- **RTDETRv2** excels in server-side processing where VRAM is abundant. Its global context awareness is perfect for [medical imaging](https://www.nature.com/subjects/medical-imaging) and dense crowd analysis where occlusions are frequent. - **DAMO-YOLO** is highly suitable for [embedded IoT applications](https://en.wikipedia.org/wiki/Internet_of_things) and fast-moving industrial inspection lines where low parameter counts and high FPS are strict requirements. ## The Future: Ultralytics YOLO26 @@ -101,7 +101,7 @@ While both RTDETRv2 and DAMO-YOLO have their merits, the field of computer visio YOLO26 adopts an **End-to-End NMS-Free Design**, capturing the primary benefit of transformers without the massive computational overhead. It incorporates the innovative **MuSGD Optimizer**—inspired by [Large Language Model](https://en.wikipedia.org/wiki/Large_language_model) training—for stable, fast convergence. Furthermore, with **DFL Removal** (Distribution Focal Loss removed for simplified export and better edge/low-power device compatibility), YOLO26 achieves up to **43% faster CPU inference**, making it the undisputed champion for [edge computing](https://en.wikipedia.org/wiki/Edge_computing). Additionally, **ProgLoss + STAL** provides improved loss functions with notable improvements in small-object recognition, critical for IoT, robotics, and aerial imagery. -Unlike models limited strictly to bounding boxes, the YOLO26 family offers unparalleled versatility, supporting tasks ranging from [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [pose estimation](https://docs.ultralytics.com/tasks/pose/) to [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), all managed seamlessly through the intuitive [Ultralytics Platform](https://platform.ultralytics.com). +Unlike models limited strictly to bounding boxes, the YOLO26 family offers unparalleled versatility, supporting tasks ranging from [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [pose estimation](https://docs.ultralytics.com/tasks/pose) to [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb), all managed seamlessly through the intuitive [Ultralytics Platform](https://platform.ultralytics.com). [Explore YOLO26 on Platform](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -123,4 +123,4 @@ Unlike models limited strictly to bounding boxes, the YOLO26 family offers unpar - **Arxiv:** [2211.15444v2](https://arxiv.org/abs/2211.15444v2) - **GitHub:** [DAMO-YOLO Repository](https://github.com/tinyvision/DAMO-YOLO) -For users interested in exploring other comparisons, check out our guides on [RTDETRv2 vs. YOLO11](https://docs.ultralytics.com/compare/rtdetr-vs-yolo11/) or [DAMO-YOLO vs. YOLOv8](https://docs.ultralytics.com/compare/damo-yolo-vs-yolov8/) to see how these models perform against previous generations of the Ultralytics family. +For users interested in exploring other comparisons, check out our guides on [RTDETRv2 vs. YOLO11](https://docs.ultralytics.com/compare/rtdetr-vs-yolo11) or [DAMO-YOLO vs. YOLOv8](https://docs.ultralytics.com/compare/damo-yolo-vs-yolov8) to see how these models perform against previous generations of the Ultralytics family. diff --git a/docs/en/compare/rtdetr-vs-efficientdet.md b/docs/en/compare/rtdetr-vs-efficientdet.md index 681a4840abd..62aa5a0057b 100644 --- a/docs/en/compare/rtdetr-vs-efficientdet.md +++ b/docs/en/compare/rtdetr-vs-efficientdet.md @@ -6,7 +6,7 @@ keywords: RTDETRv2, EfficientDet, object detection, model comparison, Vision Tra # RTDETRv2 vs. EfficientDet: Analyzing Real-Time Detection Architectures -Selecting the optimal neural network architecture is a defining choice for any [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) project. This comprehensive technical comparison dissects two influential object detection models: RTDETRv2, a state-of-the-art transformer-based detector, and EfficientDet, a highly scalable convolutional neural network. We will evaluate their distinct architectures, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/), training methodologies, and ideal deployment scenarios to help you make data-driven decisions for your AI pipelines. +Selecting the optimal neural network architecture is a defining choice for any [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) project. This comprehensive technical comparison dissects two influential object detection models: RTDETRv2, a state-of-the-art transformer-based detector, and EfficientDet, a highly scalable convolutional neural network. We will evaluate their distinct architectures, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics), training methodologies, and ideal deployment scenarios to help you make data-driven decisions for your AI pipelines. @@ -15,7 +15,7 @@ Selecting the optimal neural network architecture is a defining choice for any [ ## RTDETRv2: The Real-Time Detection Transformer -Building on the success of the original RT-DETR, RTDETRv2 refines the transformer-based [object detection](https://docs.ultralytics.com/tasks/detect/) paradigm. By optimizing the encoder and decoder structures, it delivers high accuracy while maintaining real-time inference speeds, effectively bridging the gap between traditional CNNs and vision transformers. +Building on the success of the original RT-DETR, RTDETRv2 refines the transformer-based [object detection](https://docs.ultralytics.com/tasks/detect) paradigm. By optimizing the encoder and decoder structures, it delivers high accuracy while maintaining real-time inference speeds, effectively bridging the gap between traditional CNNs and vision transformers. **Model Details** Authors: Wenyu Lv, Yian Zhao, Qinyao Chang, Kui Huang, Guanzhong Wang, and Yi Liu @@ -25,15 +25,15 @@ Links: [Arxiv](https://arxiv.org/abs/2407.17140), [GitHub](https://github.com/ly ### Architecture and Core Strengths -RTDETRv2 utilizes a hybrid architecture that pairs a potent CNN backbone (often ResNet or HGNet) with an efficient transformer decoder. The most defining characteristic of [RTDETRv2](https://docs.ultralytics.com/models/rtdetr/) is its native ability to bypass non-maximum suppression (NMS). Traditional detectors require NMS to filter out duplicate bounding boxes, adding variable [inference latency](https://www.ultralytics.com/glossary/inference-latency) during post-processing. RTDETRv2 formulates detection as a direct set prediction problem, utilizing bipartite matching to output unique predictions. +RTDETRv2 utilizes a hybrid architecture that pairs a potent CNN backbone (often ResNet or HGNet) with an efficient transformer decoder. The most defining characteristic of [RTDETRv2](https://docs.ultralytics.com/models/rtdetr) is its native ability to bypass non-maximum suppression (NMS). Traditional detectors require NMS to filter out duplicate bounding boxes, adding variable [inference latency](https://www.ultralytics.com/glossary/inference-latency) during post-processing. RTDETRv2 formulates detection as a direct set prediction problem, utilizing bipartite matching to output unique predictions. -This model excels in server-side deployments where GPU memory is abundant. Its global attention mechanism provides exceptional context awareness, making it highly adept at separating overlapping objects in dense, cluttered environments such as automated [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/) or dense crowd monitoring. +This model excels in server-side deployments where GPU memory is abundant. Its global attention mechanism provides exceptional context awareness, making it highly adept at separating overlapping objects in dense, cluttered environments such as automated [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system) or dense crowd monitoring. ### Limitations While powerful, transformer architectures inherently require more CUDA memory during training compared to standard CNNs. Furthermore, fine-tuning RTDETRv2 can require extended [training data](https://www.ultralytics.com/glossary/training-data) convergence times, making rapid prototyping slightly more resource-intensive. -[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr){ .md-button } ## EfficientDet: Scalable and Efficient CNNs @@ -53,7 +53,7 @@ EfficientDet models range from the ultra-lightweight D0 to the massive D7. This ### Limitations -EfficientDet is an older architecture that relies heavily on anchor boxes and the traditional NMS post-processing pipeline. The anchor generation process requires careful [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/), and the NMS step can bottleneck deployment on embedded hardware like a [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/). It also lacks native support for modern tasks like [pose estimation](https://docs.ultralytics.com/tasks/pose/) or [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +EfficientDet is an older architecture that relies heavily on anchor boxes and the traditional NMS post-processing pipeline. The anchor generation process requires careful [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning), and the NMS step can bottleneck deployment on embedded hardware like a [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi). It also lacks native support for modern tasks like [pose estimation](https://docs.ultralytics.com/tasks/pose) or [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb). [Learn more about EfficientDet](https://github.com/google/automl/tree/master/efficientdet){ .md-button } @@ -101,15 +101,15 @@ EfficientDet is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Alternative: Advancing the State-of-the-Art -While both RTDETRv2 and EfficientDet have strong merits, modern AI development demands frameworks that offer a seamless [developer experience](https://docs.ultralytics.com/quickstart/) alongside cutting-edge performance. The [Ultralytics ecosystem](https://docs.ultralytics.com/) provides a significantly more streamlined approach to computer vision tasks. +While both RTDETRv2 and EfficientDet have strong merits, modern AI development demands frameworks that offer a seamless [developer experience](https://docs.ultralytics.com/quickstart) alongside cutting-edge performance. The [Ultralytics ecosystem](https://docs.ultralytics.com/) provides a significantly more streamlined approach to computer vision tasks. If you are exploring state-of-the-art detection, the newly released [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) synthesizes the best aspects of both CNNs and transformers. @@ -117,7 +117,7 @@ If you are exploring state-of-the-art detection, the newly released [Ultralytics YOLO26 implements an **End-to-End NMS-Free Design**, bringing the deployment simplicity of RTDETRv2 to the ultra-efficient YOLO architecture. Furthermore, it introduces the **MuSGD Optimizer**—inspired by LLM training innovations—for superior training stability. With **DFL Removal** (Distribution Focal Loss removed for simplified export and better edge/low-power device compatibility), YOLO26 boasts up to **43% faster CPU inference** than previous generations, making it an exceptional choice for [edge computing](https://www.ultralytics.com/glossary/edge-computing) over heavier models. Additionally, **ProgLoss + STAL** delivers improved loss functions with notable improvements in small-object recognition, critical for IoT, robotics, and aerial imagery. -The ease of use provided by the [Ultralytics Python package](https://docs.ultralytics.com/usage/python/) is unmatched. Developers can train, validate, and [export models](https://docs.ultralytics.com/modes/export/) using an intuitive API that abstracts away the boilerplate code typically required by research repositories. +The ease of use provided by the [Ultralytics Python package](https://docs.ultralytics.com/usage/python) is unmatched. Developers can train, validate, and [export models](https://docs.ultralytics.com/modes/export) using an intuitive API that abstracts away the boilerplate code typically required by research repositories. ```python from ultralytics import RTDETR @@ -132,6 +132,6 @@ results = model.train(data="coco8.yaml", epochs=100, imgsz=640) model.export(format="engine") ``` -Ultralytics models natively support multiple tasks, including [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [image classification](https://docs.ultralytics.com/tasks/classify/), providing a versatile toolkit for diverse industry needs. Furthermore, the removal of Distribution Focal Loss (DFL) in modern Ultralytics models simplifies the computational graph, guaranteeing smoother export to embedded [NPUs and TPUs](https://docs.ultralytics.com/integrations/edge-tpu/). +Ultralytics models natively support multiple tasks, including [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [image classification](https://docs.ultralytics.com/tasks/classify), providing a versatile toolkit for diverse industry needs. Furthermore, the removal of Distribution Focal Loss (DFL) in modern Ultralytics models simplifies the computational graph, guaranteeing smoother export to embedded [NPUs and TPUs](https://docs.ultralytics.com/integrations/edge-tpu). -For seamless [data annotation](https://docs.ultralytics.com/platform/data/annotation/) and model management, the [Ultralytics Platform](https://platform.ultralytics.com/) provides a comprehensive cloud environment to oversee the entire machine learning lifecycle, establishing it as the premier choice for deploying robust computer vision solutions in production. +For seamless [data annotation](https://docs.ultralytics.com/platform/data/annotation) and model management, the [Ultralytics Platform](https://platform.ultralytics.com/) provides a comprehensive cloud environment to oversee the entire machine learning lifecycle, establishing it as the premier choice for deploying robust computer vision solutions in production. diff --git a/docs/en/compare/rtdetr-vs-pp-yoloe.md b/docs/en/compare/rtdetr-vs-pp-yoloe.md index 3ce10e31a93..d55c93a31f5 100644 --- a/docs/en/compare/rtdetr-vs-pp-yoloe.md +++ b/docs/en/compare/rtdetr-vs-pp-yoloe.md @@ -6,7 +6,7 @@ keywords: RTDETRv2,PP-YOLOE+,object detection,model comparison,Vision Transforme # RTDETRv2 vs. PP-YOLOE+: A Technical Comparison of Object Detection Models -The rapidly evolving field of computer vision has produced diverse architectural approaches to solve complex [real-time object detection](https://docs.ultralytics.com/tasks/detect/) challenges. Among the most notable recent advancements are **RTDETRv2** and **PP-YOLOE+**, two powerful models that approach visual recognition from fundamentally different design philosophies. While both models aim to provide high-performance detection, their underlying mechanics, training paradigms, and ideal deployment scenarios vary significantly. +The rapidly evolving field of computer vision has produced diverse architectural approaches to solve complex [real-time object detection](https://docs.ultralytics.com/tasks/detect) challenges. Among the most notable recent advancements are **RTDETRv2** and **PP-YOLOE+**, two powerful models that approach visual recognition from fundamentally different design philosophies. While both models aim to provide high-performance detection, their underlying mechanics, training paradigms, and ideal deployment scenarios vary significantly. This comprehensive guide delves into the technical nuances of both models, comparing their architectures, performance metrics, and ecosystem support to help developers and researchers choose the optimal solution for their specific deployment needs. @@ -29,7 +29,7 @@ Date: 2024-07-24 Arxiv: [2407.17140](https://arxiv.org/abs/2407.17140) GitHub: [RT-DETR Repository](https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetrv2_pytorch) -[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr){ .md-button } ### PP-YOLOE+ @@ -45,7 +45,7 @@ GitHub: [PaddleDetection Repository](https://github.com/PaddlePaddle/PaddleDetec !!! tip "Ecosystem Integration" - While both models have their standalone research repositories, you can easily experiment with RTDETRv2 directly within the [Ultralytics Python package](https://docs.ultralytics.com/usage/python/), benefiting from a unified API and streamlined export options. + While both models have their standalone research repositories, you can easily experiment with RTDETRv2 directly within the [Ultralytics Python package](https://docs.ultralytics.com/usage/python), benefiting from a unified API and streamlined export options. ## Architectural Differences @@ -57,7 +57,7 @@ Conversely, RTDETRv2 employs a Hybrid Encoder and a Transformer Decoder. This al ## Performance Metrics and Comparison -When evaluating [YOLO performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/), it is crucial to balance accuracy (mAP) against computational cost (FLOPs) and inference speed. The table below highlights the performance of both models across various sizes. +When evaluating [YOLO performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics), it is crucial to balance accuracy (mAP) against computational cost (FLOPs) and inference speed. The table below highlights the performance of both models across various sizes. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -84,13 +84,13 @@ YOLO26 synthesizes the best aspects of both CNNs and Transformers. It adopts the Unlike heavy transformer models that demand substantial CUDA memory, YOLO26 features **DFL Removal** (Distribution Focal Loss) and is specifically optimized for edge computing, delivering up to **43% faster CPU inference** compared to previous generations. -Additionally, YOLO26 is not limited to simple object detection. It is natively versatile, supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) out of the box, whereas PP-YOLOE+ is primarily focused on bounding box detection. +Additionally, YOLO26 is not limited to simple object detection. It is natively versatile, supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb) out of the box, whereas PP-YOLOE+ is primarily focused on bounding box detection. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } ## Training Methodologies and Ecosystem -Training efficiency and ease of use are where the [Ultralytics ecosystem](https://docs.ultralytics.com/platform/) truly shines compared to standalone research repositories. While PP-YOLOE+ relies on the PaddlePaddle framework and RTDETRv2 often requires complex environment setups, integrating models through Ultralytics provides a seamless experience. +Training efficiency and ease of use are where the [Ultralytics ecosystem](https://docs.ultralytics.com/platform) truly shines compared to standalone research repositories. While PP-YOLOE+ relies on the PaddlePaddle framework and RTDETRv2 often requires complex environment setups, integrating models through Ultralytics provides a seamless experience. With the Ultralytics API, you benefit from lower memory requirements during training, automated dataset handling, and simplified hyperparameter tuning. Furthermore, deploying models to production formats like [ONNX](https://onnxruntime.ai/) or [TensorRT](https://developer.nvidia.com/tensorrt) can be accomplished with a single command. @@ -148,11 +148,11 @@ PP-YOLOE+ is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Conclusion diff --git a/docs/en/compare/rtdetr-vs-yolo11.md b/docs/en/compare/rtdetr-vs-yolo11.md index f6bf4fd6c05..1eff0625bd8 100644 --- a/docs/en/compare/rtdetr-vs-yolo11.md +++ b/docs/en/compare/rtdetr-vs-yolo11.md @@ -30,11 +30,11 @@ Introduced as an evolution of the original Real-Time Detection Transformer, RTDE ### Architectural Strengths and Weaknesses -RTDETRv2's primary innovation is its end-to-end NMS-free architecture. By eliminating Non-Maximum Suppression (NMS), it simplifies the post-processing pipeline. Furthermore, its multi-scale feature extraction capabilities have been improved over the original [RT-DETR model](https://docs.ultralytics.com/models/rtdetr/), allowing it to better identify objects of varying sizes. +RTDETRv2's primary innovation is its end-to-end NMS-free architecture. By eliminating Non-Maximum Suppression (NMS), it simplifies the post-processing pipeline. Furthermore, its multi-scale feature extraction capabilities have been improved over the original [RT-DETR model](https://docs.ultralytics.com/models/rtdetr), allowing it to better identify objects of varying sizes. However, because it relies on Transformers, RTDETRv2 typically suffers from significantly higher memory requirements during training. Transformers are generally slower to converge and require substantially more CUDA memory compared to traditional CNNs, making them less accessible for researchers operating on consumer-grade hardware or deploying to constrained [edge AI](https://www.ultralytics.com/glossary/edge-ai) environments. -[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr){ .md-button } ## Ultralytics YOLO11: The Pinnacle of CNN Efficiency @@ -51,7 +51,7 @@ Building upon years of foundational research, Ultralytics released YOLO11 as a m YOLO11 shines in its **Performance Balance**. It achieves an extraordinary trade-off between speed and accuracy, making it exceptionally versatile for diverse real-world deployment scenarios, from massive [cloud computing](https://www.ultralytics.com/glossary/cloud-computing) clusters to lightweight mobile devices. -Moreover, Ultralytics YOLO models are renowned for their lower memory usage during training and inference. Unlike Transformer models which can easily exhaust VRAM, YOLO11 allows for larger batch sizes on standard GPUs. Furthermore, YOLO11 is not limited to mere object detection; it boasts incredible **Versatility**, featuring native support for [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +Moreover, Ultralytics YOLO models are renowned for their lower memory usage during training and inference. Unlike Transformer models which can easily exhaust VRAM, YOLO11 allows for larger batch sizes on standard GPUs. Furthermore, YOLO11 is not limited to mere object detection; it boasts incredible **Versatility**, featuring native support for [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Image Classification](https://docs.ultralytics.com/tasks/classify), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } @@ -72,13 +72,13 @@ When comparing raw numbers, it becomes evident that while RTDETRv2 achieves impr | YOLO11l | 640 | 53.4 | 238.6 | 6.2 | 25.3 | 86.9 | | YOLO11x | 640 | **54.7** | 462.8 | 11.3 | 56.9 | 194.9 | -As seen in the table, the **YOLO11x** model achieves a superior mAPval of 54.7% while utilizing fewer FLOPs (194.9B vs 259B) and delivering faster inference on TensorRT (11.3ms vs 15.03ms) compared to the RTDETRv2-x variant. The nano and small YOLO11 variants provide unparalleled lightweight options for constrained devices like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/). +As seen in the table, the **YOLO11x** model achieves a superior mAPval of 54.7% while utilizing fewer FLOPs (194.9B vs 259B) and delivering faster inference on TensorRT (11.3ms vs 15.03ms) compared to the RTDETRv2-x variant. The nano and small YOLO11 variants provide unparalleled lightweight options for constrained devices like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi). ## Ecosystem, Ease of Use, and Training -The defining characteristic of Ultralytics models is the streamlined user experience. The `ultralytics` Python package provides a unified, intuitive API that handles the heavy lifting of [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation/), distributed training, and model export. While RTDETRv2's research repository requires significant boilerplate and configuration, Ultralytics provides a "zero-to-hero" pipeline. +The defining characteristic of Ultralytics models is the streamlined user experience. The `ultralytics` Python package provides a unified, intuitive API that handles the heavy lifting of [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation), distributed training, and model export. While RTDETRv2's research repository requires significant boilerplate and configuration, Ultralytics provides a "zero-to-hero" pipeline. -Interestingly, the Ultralytics ecosystem is so robust that it natively supports running RT-DETR models alongside YOLO models! This allows you to leverage the **Well-Maintained Ecosystem** of Ultralytics—including integrations with [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) and [Comet ML](https://docs.ultralytics.com/integrations/comet/)—for tracking experiments effortlessly. +Interestingly, the Ultralytics ecosystem is so robust that it natively supports running RT-DETR models alongside YOLO models! This allows you to leverage the **Well-Maintained Ecosystem** of Ultralytics—including integrations with [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) and [Comet ML](https://docs.ultralytics.com/integrations/comet)—for tracking experiments effortlessly. ```python from ultralytics import RTDETR, YOLO @@ -130,17 +130,17 @@ RT-DETR is a strong choice for: YOLO11 is recommended for: -- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) where reliability and active maintenance are paramount. -- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [OBB](https://docs.ultralytics.com/tasks/obb/) within a single unified framework. -- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python/). +- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) where reliability and active maintenance are paramount. +- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [OBB](https://docs.ultralytics.com/tasks/obb) within a single unified framework. +- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python). ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Looking Forward: The Arrival of YOLO26 diff --git a/docs/en/compare/rtdetr-vs-yolo26.md b/docs/en/compare/rtdetr-vs-yolo26.md index b73ffbb4b4b..dae4955d063 100644 --- a/docs/en/compare/rtdetr-vs-yolo26.md +++ b/docs/en/compare/rtdetr-vs-yolo26.md @@ -15,7 +15,7 @@ The landscape of real-time object detection has evolved dramatically, with resea ## RTDETRv2: Real-Time Detection Transformers -RTDETRv2 builds upon the original [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) architecture, aiming to combine the global context awareness of vision transformers with the speed required for real-time applications. +RTDETRv2 builds upon the original [RT-DETR](https://docs.ultralytics.com/models/rtdetr) architecture, aiming to combine the global context awareness of vision transformers with the speed required for real-time applications. **Key Characteristics:** @@ -26,7 +26,7 @@ RTDETRv2 builds upon the original [RT-DETR](https://docs.ultralytics.com/models/ ### Architecture and Strengths -Unlike traditional anchor-based detectors, RTDETRv2 leverages a transformer-based approach that natively eliminates the need for [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) during post-processing. By utilizing a flexible attention mechanism, the model is highly effective at understanding complex scenes and overlapping objects. Its "Bag-of-Freebies" improvements have significantly enhanced its accuracy on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/) while maintaining acceptable inference speeds on high-end GPUs. +Unlike traditional anchor-based detectors, RTDETRv2 leverages a transformer-based approach that natively eliminates the need for [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) during post-processing. By utilizing a flexible attention mechanism, the model is highly effective at understanding complex scenes and overlapping objects. Its "Bag-of-Freebies" improvements have significantly enhanced its accuracy on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco) while maintaining acceptable inference speeds on high-end GPUs. ### Limitations @@ -43,20 +43,20 @@ Released in early 2026, **Ultralytics YOLO26** redefines what is possible with C - **Authors:** Glenn Jocher and Jing Qiu - **Organization:** [Ultralytics](https://www.ultralytics.com/about) - **Date:** January 14, 2026 -- **Links:** [GitHub](https://github.com/ultralytics/ultralytics), [Docs](https://docs.ultralytics.com/models/yolo26/) +- **Links:** [GitHub](https://github.com/ultralytics/ultralytics), [Docs](https://docs.ultralytics.com/models/yolo26) ### Architectural Breakthroughs YOLO26 introduces several revolutionary features that solve common pain points in model deployment: -- **End-to-End NMS-Free Design:** Building on concepts pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 is natively end-to-end. By removing NMS post-processing, it drastically reduces latency variability, ensuring highly predictable inference times in production. +- **End-to-End NMS-Free Design:** Building on concepts pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 is natively end-to-end. By removing NMS post-processing, it drastically reduces latency variability, ensuring highly predictable inference times in production. - **Up to 43% Faster CPU Inference:** Through strategic architectural refinements and the removal of Distribution Focal Loss (DFL), YOLO26 achieves unprecedented CPU speeds, making it the premier choice for [edge computing](https://www.ultralytics.com/glossary/edge-computing) without dedicated GPUs. - **MuSGD Optimizer:** Inspired by Large Language Model (LLM) training techniques like Moonshot AI's Kimi K2, YOLO26 utilizes the MuSGD optimizer (a hybrid of SGD and Muon). This ensures highly stable training runs and incredibly fast convergence. - **ProgLoss + STAL:** These advanced loss functions deliver remarkable improvements in small-object recognition, an essential upgrade for applications involving [aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision) and drone-based surveillance. !!! tip "Task-Specific Enhancements in YOLO26" - Beyond standard detection, YOLO26 features specialized improvements: Semantic segmentation loss and multi-scale proto for [segmentation tasks](https://docs.ultralytics.com/tasks/segment/), Residual Log-Likelihood Estimation (RLE) for [pose estimation](https://docs.ultralytics.com/tasks/pose/), and customized angle loss to resolve boundary issues in [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. + Beyond standard detection, YOLO26 features specialized improvements: Semantic segmentation loss and multi-scale proto for [segmentation tasks](https://docs.ultralytics.com/tasks/segment), Residual Log-Likelihood Estimation (RLE) for [pose estimation](https://docs.ultralytics.com/tasks/pose), and customized angle loss to resolve boundary issues in [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -77,7 +77,7 @@ When evaluating these models, achieving a strong performance balance between acc | YOLO26l | 640 | 55.0 | 286.2 | 6.2 | 24.8 | 86.4 | | YOLO26x | 640 | **57.5** | 525.8 | 11.8 | 55.7 | 193.9 | -As seen above, the YOLO26x model achieves a remarkable **57.5 mAP**, significantly surpassing the RTDETRv2-x model while utilizing fewer parameters and maintaining a faster [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) inference speed. Furthermore, the memory requirements for YOLO26 are noticeably lower, making it the optimal choice for real-time edge deployments. +As seen above, the YOLO26x model achieves a remarkable **57.5 mAP**, significantly surpassing the RTDETRv2-x model while utilizing fewer parameters and maintaining a faster [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) inference speed. Furthermore, the memory requirements for YOLO26 are noticeably lower, making it the optimal choice for real-time edge deployments. ## Ecosystem and Ease of Use @@ -87,7 +87,7 @@ While raw performance is vital, the surrounding ecosystem dictates how quickly a RTDETRv2 operates primarily as a research-grade repository, which can necessitate complex environment setups and manual scripting for custom tasks. Conversely, Ultralytics YOLO26 benefits from a mature, heavily tested Python package. The Ultralytics ecosystem provides an incredibly streamlined user experience, offering a simple API for training, validation, prediction, and export. -With built-in integrations for [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) and [Comet ML](https://docs.ultralytics.com/integrations/comet/), experiment tracking is seamless. Furthermore, Ultralytics models are highly versatile; while RTDETRv2 focuses on object detection, YOLO26 natively supports instance segmentation, pose estimation, and image classification within the exact same framework. +With built-in integrations for [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) and [Comet ML](https://docs.ultralytics.com/integrations/comet), experiment tracking is seamless. Furthermore, Ultralytics models are highly versatile; while RTDETRv2 focuses on object detection, YOLO26 natively supports instance segmentation, pose estimation, and image classification within the exact same framework. ### Code Example: Simplicity in Action @@ -131,12 +131,12 @@ YOLO26 is recommended for: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Exploring Other Architectures -While YOLO26 represents the current pinnacle of performance, developers might also find value in exploring previous iterations. The highly successful [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a robust, fully supported model for a variety of legacy systems. You can dive deeper into its capabilities by reading our [RTDETR vs YOLO11 comparison](https://docs.ultralytics.com/compare/rtdetr-vs-yolo11/). Additionally, if you are analyzing older architectures, checking out the [EfficientDet vs YOLO26 comparison](https://docs.ultralytics.com/compare/efficientdet-vs-yolo26/) provides great historical context on how far [object detection architectures](https://www.ultralytics.com/glossary/object-detection-architectures) have progressed. +While YOLO26 represents the current pinnacle of performance, developers might also find value in exploring previous iterations. The highly successful [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a robust, fully supported model for a variety of legacy systems. You can dive deeper into its capabilities by reading our [RTDETR vs YOLO11 comparison](https://docs.ultralytics.com/compare/rtdetr-vs-yolo11). Additionally, if you are analyzing older architectures, checking out the [EfficientDet vs YOLO26 comparison](https://docs.ultralytics.com/compare/efficientdet-vs-yolo26) provides great historical context on how far [object detection architectures](https://www.ultralytics.com/glossary/object-detection-architectures) have progressed. ## Final Thoughts -Both RTDETRv2 and YOLO26 offer incredible advancements in the field of AI. However, for teams prioritizing a seamless transition to production, minimal memory footprint, and broad task versatility, **Ultralytics YOLO26** is the clear recommendation. Its NMS-free architecture, rapid CPU speeds, and the backing of the robust Ultralytics ecosystem ensure that your vision AI projects remain scalable, efficient, and future-proof. Whether deploying on a cloud server or a resource-limited [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/), YOLO26 delivers uncompromising performance out of the box. +Both RTDETRv2 and YOLO26 offer incredible advancements in the field of AI. However, for teams prioritizing a seamless transition to production, minimal memory footprint, and broad task versatility, **Ultralytics YOLO26** is the clear recommendation. Its NMS-free architecture, rapid CPU speeds, and the backing of the robust Ultralytics ecosystem ensure that your vision AI projects remain scalable, efficient, and future-proof. Whether deploying on a cloud server or a resource-limited [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi), YOLO26 delivers uncompromising performance out of the box. diff --git a/docs/en/compare/rtdetr-vs-yolov10.md b/docs/en/compare/rtdetr-vs-yolov10.md index c950b4396dc..38c3896d9d2 100644 --- a/docs/en/compare/rtdetr-vs-yolov10.md +++ b/docs/en/compare/rtdetr-vs-yolov10.md @@ -17,7 +17,7 @@ This guide provides a comprehensive technical comparison of these two models, an ## RTDETRv2: Real-Time Detection Transformers -RTDETRv2 builds upon the original [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) architecture, aiming to combine the global context understanding of Vision Transformers with the real-time speed requirements traditionally dominated by YOLO models. +RTDETRv2 builds upon the original [RT-DETR](https://docs.ultralytics.com/models/rtdetr) architecture, aiming to combine the global context understanding of Vision Transformers with the real-time speed requirements traditionally dominated by YOLO models. **Key Characteristics:** @@ -37,7 +37,7 @@ RTDETRv2 utilizes an end-to-end transformer architecture that inherently avoids - **Global Context:** The [attention mechanism](https://www.ultralytics.com/glossary/attention-mechanism) allows the model to excel in complex, cluttered environments. - **NMS-Free:** Directly predicts object coordinates, simplifying the deployment pipeline. -- **High Accuracy:** Achieves excellent [mean average precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics/) on the COCO dataset. +- **High Accuracy:** Achieves excellent [mean average precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics) on the COCO dataset. **Weaknesses:** @@ -72,9 +72,9 @@ The core innovation of YOLOv10 is its consistent dual assignments for NMS-free t **Weaknesses:** -- **First-Generation Concept:** As the first YOLO to implement this specific NMS-free architecture, it laid the groundwork but left room for the multi-task versatility and optimization seen in subsequent models like [YOLO11](https://docs.ultralytics.com/models/yolo11/) and YOLO26. +- **First-Generation Concept:** As the first YOLO to implement this specific NMS-free architecture, it laid the groundwork but left room for the multi-task versatility and optimization seen in subsequent models like [YOLO11](https://docs.ultralytics.com/models/yolo11) and YOLO26. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## Performance Comparison @@ -122,11 +122,11 @@ YOLOv10 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Introducing YOLO26 @@ -139,13 +139,13 @@ For developers seeking the absolute state-of-the-art in 2026, **[Ultralytics YOL - **MuSGD Optimizer:** A hybrid of SGD and Muon (inspired by LLM training innovations), this novel optimizer provides more stable training and significantly faster convergence compared to traditional methods. - **Up to 43% Faster CPU Inference:** Carefully optimized for environments without dedicated GPUs, democratizing high-performance vision AI. - **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, which is critical for [applications using drones](https://www.ultralytics.com/blog/computer-vision-applications-ai-drone-uav-operations) and IoT sensors. -- **Unmatched Versatility:** Unlike models limited to bounding boxes, YOLO26 supports a full suite of tasks including [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [OBB detection](https://docs.ultralytics.com/tasks/obb/), complete with task-specific improvements like Residual Log-Likelihood Estimation (RLE) for Pose. +- **Unmatched Versatility:** Unlike models limited to bounding boxes, YOLO26 supports a full suite of tasks including [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [OBB detection](https://docs.ultralytics.com/tasks/obb), complete with task-specific improvements like Residual Log-Likelihood Estimation (RLE) for Pose. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } ### Seamless Implementation with Python -Training and deploying these models using the [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) is designed to be frictionless. Memory requirements are notably lower during training compared to transformer-heavy architectures, allowing you to train powerful models on standard hardware. +Training and deploying these models using the [Ultralytics Python API](https://docs.ultralytics.com/usage/python) is designed to be frictionless. Memory requirements are notably lower during training compared to transformer-heavy architectures, allowing you to train powerful models on standard hardware. ```python from ultralytics import YOLO @@ -161,4 +161,4 @@ results = model.train(data="coco8.yaml", epochs=100, imgsz=640) model.export(format="onnx", simplify=True) ``` -Whether you are implementing [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/) or conducting [medical image analysis](https://www.ultralytics.com/glossary/medical-image-analysis), choosing a model backed by the active Ultralytics community ensures you have the tools, [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/) guides, and continuous updates needed to succeed. While YOLOv10 and RTDETRv2 paved the way for NMS-free architectures, YOLO26 perfects the formula, offering the best balance of performance, versatility, and production readiness. +Whether you are implementing [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system) or conducting [medical image analysis](https://www.ultralytics.com/glossary/medical-image-analysis), choosing a model backed by the active Ultralytics community ensures you have the tools, [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning) guides, and continuous updates needed to succeed. While YOLOv10 and RTDETRv2 paved the way for NMS-free architectures, YOLO26 perfects the formula, offering the best balance of performance, versatility, and production readiness. diff --git a/docs/en/compare/rtdetr-vs-yolov5.md b/docs/en/compare/rtdetr-vs-yolov5.md index 93be9637205..43f354f15b7 100644 --- a/docs/en/compare/rtdetr-vs-yolov5.md +++ b/docs/en/compare/rtdetr-vs-yolov5.md @@ -8,7 +8,7 @@ keywords: YOLOv5, RTDETRv2, object detection comparison, YOLOv5 vs RTDETRv2, Ult The evolution of [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) has been largely defined by the relentless pursuit of balancing accuracy with real-time inference speed. When comparing RTDETRv2 and Ultralytics YOLOv5, developers are essentially weighing the sophisticated global context capabilities of transformer architectures against the highly optimized, battle-tested efficiency of Convolutional Neural Networks (CNNs). -This guide provides an in-depth technical analysis of these two prominent architectures, detailing their performance metrics, training methodologies, memory requirements, and ideal deployment scenarios to help you choose the best [object detection](https://docs.ultralytics.com/tasks/detect/) model for your specific use case. +This guide provides an in-depth technical analysis of these two prominent architectures, detailing their performance metrics, training methodologies, memory requirements, and ideal deployment scenarios to help you choose the best [object detection](https://docs.ultralytics.com/tasks/detect) model for your specific use case. @@ -39,13 +39,13 @@ Ultralytics YOLOv5 fundamentally changed the landscape of applied machine learni - **Author:** Glenn Jocher - **Organization:** Ultralytics - **Date:** June 26, 2020 -- **Links:** [Official Documentation](https://docs.ultralytics.com/models/yolov5/), [GitHub Repository](https://github.com/ultralytics/yolov5) +- **Links:** [Official Documentation](https://docs.ultralytics.com/models/yolov5), [GitHub Repository](https://github.com/ultralytics/yolov5) ### Ecosystem and Performance Balance YOLOv5 is built entirely on the [PyTorch](https://pytorch.org/) framework and relies on an immensely efficient CNN architecture. It was designed from the ground up for **ease of use**, featuring a streamlined API and some of the most extensive documentation in the AI industry. -The greatest advantage of YOLOv5 lies in its unmatched versatility and low memory requirements. Training a YOLOv5 model requires drastically less VRAM than transformer-based models, making it accessible to researchers and engineers with limited hardware budgets. Furthermore, while RTDETRv2 focuses exclusively on bounding box detection, YOLOv5 has evolved into a versatile powerhouse supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [image classification](https://docs.ultralytics.com/tasks/classify/). +The greatest advantage of YOLOv5 lies in its unmatched versatility and low memory requirements. Training a YOLOv5 model requires drastically less VRAM than transformer-based models, making it accessible to researchers and engineers with limited hardware budgets. Furthermore, while RTDETRv2 focuses exclusively on bounding box detection, YOLOv5 has evolved into a versatile powerhouse supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [image classification](https://docs.ultralytics.com/tasks/classify). !!! tip "Enterprise Model Management" @@ -55,7 +55,7 @@ The greatest advantage of YOLOv5 lies in its unmatched versatility and low memor ## Performance and Metrics Comparison -When analyzing raw performance on the standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/), we can see clear distinctions in how these models prioritize resources. +When analyzing raw performance on the standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco), we can see clear distinctions in how these models prioritize resources. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -97,7 +97,7 @@ results_rtdetr = model_rtdetr("https://ultralytics.com/images/bus.jpg") results_yolo[0].show() ``` -By leveraging the Ultralytics library, developers automatically gain access to a well-maintained ecosystem featuring [experiment tracking integrations](https://docs.ultralytics.com/integrations/weights-biases/) (like Weights & Biases and Comet ML) and one-click exports to deployment formats like [ONNX](https://onnx.ai/) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino/). +By leveraging the Ultralytics library, developers automatically gain access to a well-maintained ecosystem featuring [experiment tracking integrations](https://docs.ultralytics.com/integrations/weights-biases) (like Weights & Biases and Comet ML) and one-click exports to deployment formats like [ONNX](https://onnx.ai/) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino). ## Real-World Applications and Ideal Use Cases @@ -112,12 +112,12 @@ RTDETRv2 is best suited for environments where hardware limitations are non-exis YOLOv5 is the undeniable champion for practical, real-world deployment across diverse hardware. -- **Edge AI Devices:** Deploying [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/) on Raspberry Pi or NVIDIA Jetson devices where memory is strictly limited. +- **Edge AI Devices:** Deploying [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system) on Raspberry Pi or NVIDIA Jetson devices where memory is strictly limited. - **Mobile Applications:** Running fast, real-time bounding box and segmentation inference directly on smartphones via CoreML or TFLite. - **High-Speed Industrial Manufacturing:** Inspecting parts on rapid production lines where millisecond latency is critical to operational success. !!! note "Exploring Other Ultralytics Models" - While YOLOv5 is a legendary model, the Ultralytics ecosystem continually pushes the boundaries of AI. If you are comparing models for a new project in 2026, you should consider exploring the state-of-the-art [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26). YOLO26 incorporates a native **End-to-End NMS-Free Design** (similar to transformers but with CNN speed), features the revolutionary **MuSGD Optimizer** for incredibly stable training, and delivers up to 43% faster CPU inference. Alternatively, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a fantastic, highly supported choice for versatile deployments requiring [Pose Estimation](https://docs.ultralytics.com/tasks/pose/) and [OBB detection](https://docs.ultralytics.com/tasks/obb/). + While YOLOv5 is a legendary model, the Ultralytics ecosystem continually pushes the boundaries of AI. If you are comparing models for a new project in 2026, you should consider exploring the state-of-the-art [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26). YOLO26 incorporates a native **End-to-End NMS-Free Design** (similar to transformers but with CNN speed), features the revolutionary **MuSGD Optimizer** for incredibly stable training, and delivers up to 43% faster CPU inference. Alternatively, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a fantastic, highly supported choice for versatile deployments requiring [Pose Estimation](https://docs.ultralytics.com/tasks/pose) and [OBB detection](https://docs.ultralytics.com/tasks/obb). Ultimately, while RTDETRv2 pushes the accuracy ceiling using transformer layers, the Ultralytics YOLO framework provides an unmatched balance of speed, lightweight memory requirements, and a brilliantly engineered developer experience that dramatically reduces the time from prototype to production. diff --git a/docs/en/compare/rtdetr-vs-yolov6.md b/docs/en/compare/rtdetr-vs-yolov6.md index 8b260e67aea..8dee7e47913 100644 --- a/docs/en/compare/rtdetr-vs-yolov6.md +++ b/docs/en/compare/rtdetr-vs-yolov6.md @@ -17,7 +17,7 @@ This comprehensive technical comparison explores their respective architectures, ## RTDETRv2: The Vision Transformer Approach -Developed by researchers at Baidu, RTDETRv2 builds upon the foundation of the original RT-DETR, representing a significant leap forward in transformer-based [object detection](https://docs.ultralytics.com/tasks/detect/). +Developed by researchers at Baidu, RTDETRv2 builds upon the foundation of the original RT-DETR, representing a significant leap forward in transformer-based [object detection](https://docs.ultralytics.com/tasks/detect). - Authors: Wenyu Lv, Yian Zhao, Qinyao Chang, Kui Huang, Guanzhong Wang, and Yi Liu - Organization: [Baidu](https://www.baidu.com/) @@ -36,7 +36,7 @@ The "Bag-of-Freebies" incorporated into RTDETRv2 enhances its ability to handle While transformers excel at complex scene understanding, they typically require significantly higher CUDA memory during training compared to CNNs. This can limit batch sizes on standard consumer GPUs and increase overall training time. -[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr){ .md-button } ## YOLOv6-3.0: Industrial Throughput Maximization @@ -54,7 +54,7 @@ YOLOv6-3.0 relies on an **EfficientRep** backbone, meticulously designed to mini During training, it employs an Anchor-Aided Training (AAT) strategy to benefit from anchor-based paradigms while maintaining an anchor-free inference mode for faster execution. While it achieves exceptional throughput on server-grade GPUs (e.g., T4, A100), its specialized architecture can result in suboptimal latency when deployed on CPU-only edge devices. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Performance Comparison @@ -72,7 +72,7 @@ When evaluating models for production, balancing accuracy (mAP) with inference s | YOLOv6-3.0m | 640 | 50.0 | - | 5.28 | 34.9 | 85.8 | | YOLOv6-3.0l | 640 | 52.8 | - | 8.95 | 59.6 | 150.7 | -While YOLOv6-3.0 dominates in sheer processing speed on TensorRT, RTDETRv2 captures higher mAP scores, particularly scaling better with larger model variants. However, both models lack the extensive versatility found in modern unified frameworks. YOLOv6-3.0 is primarily a detection specialist, missing native support for tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [pose estimation](https://docs.ultralytics.com/tasks/pose/) out of the box. +While YOLOv6-3.0 dominates in sheer processing speed on TensorRT, RTDETRv2 captures higher mAP scores, particularly scaling better with larger model variants. However, both models lack the extensive versatility found in modern unified frameworks. YOLOv6-3.0 is primarily a detection specialist, missing native support for tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [pose estimation](https://docs.ultralytics.com/tasks/pose) out of the box. ## Use Cases and Recommendations @@ -96,20 +96,20 @@ YOLOv6 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage Choosing the right model involves more than just raw benchmark numbers; developer experience, deployment flexibility, and ecosystem support are equally crucial. By utilizing models integrated within the Ultralytics platform, users gain significant advantages over static research repositories. - **Ease of Use:** The `ultralytics` Python package offers a seamless API. Training, validating, and exporting models takes only a few lines of code. -- **Well-Maintained Ecosystem:** Unlike isolated academic repos, the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolov8) is actively updated. It boasts robust integrations for tools like [ONNX](https://docs.ultralytics.com/integrations/onnx/), [OpenVINO](https://docs.ultralytics.com/integrations/openvino/), and CoreML. +- **Well-Maintained Ecosystem:** Unlike isolated academic repos, the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolov8) is actively updated. It boasts robust integrations for tools like [ONNX](https://docs.ultralytics.com/integrations/onnx), [OpenVINO](https://docs.ultralytics.com/integrations/openvino), and CoreML. - **Training Efficiency:** Ultralytics models typically consume significantly lower VRAM during training compared to transformer architectures like RTDETRv2, allowing for larger batch sizes on consumer-grade hardware. -- **Versatility:** Unlike the focused scope of YOLOv6-3.0, Ultralytics models are multi-modal, natively supporting [image classification](https://docs.ultralytics.com/tasks/classify/), [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), and segmentation within a single unified framework. +- **Versatility:** Unlike the focused scope of YOLOv6-3.0, Ultralytics models are multi-modal, natively supporting [image classification](https://docs.ultralytics.com/tasks/classify), [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb), and segmentation within a single unified framework. !!! tip "Streamlined Deployment" @@ -121,14 +121,14 @@ While RTDETRv2 and YOLOv6-3.0 offer specific benefits, the field moves rapidly. YOLO26 synthesizes the strengths of industrial CNNs and modern transformers while eliminating their respective weaknesses: -- **End-to-End NMS-Free Design:** Adopting the breakthrough first introduced in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 eliminates NMS post-processing natively, ensuring stable, predictable deployment similar to RTDETRv2 but with far less overhead. +- **End-to-End NMS-Free Design:** Adopting the breakthrough first introduced in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 eliminates NMS post-processing natively, ensuring stable, predictable deployment similar to RTDETRv2 but with far less overhead. - **MuSGD Optimizer:** Inspired by advanced LLM training techniques (such as Moonshot AI's Kimi K2), this hybrid optimizer ensures stable training and faster convergence, overcoming the notorious instability of traditional vision transformers. - **Optimized for Edge:** With up to **43% faster CPU inference** than previous generations and the strategic removal of Distribution Focal Loss (DFL), YOLO26 is perfectly suited for mobile and IoT devices where GPU acceleration isn't available. - **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, a historic challenge for CNNs, making YOLO26 ideal for aerial imagery and robotics. ### Training Example -The intuitive Ultralytics API allows you to train cutting-edge models seamlessly. Below is a runnable example demonstrating how to train the YOLO26 Nano model on the [COCO8 dataset](https://docs.ultralytics.com/datasets/detect/coco8/): +The intuitive Ultralytics API allows you to train cutting-edge models seamlessly. Below is a runnable example demonstrating how to train the YOLO26 Nano model on the [COCO8 dataset](https://docs.ultralytics.com/datasets/detect/coco8): ```python from ultralytics import YOLO diff --git a/docs/en/compare/rtdetr-vs-yolov7.md b/docs/en/compare/rtdetr-vs-yolov7.md index 4a0ce672330..e5ae9431a2b 100644 --- a/docs/en/compare/rtdetr-vs-yolov7.md +++ b/docs/en/compare/rtdetr-vs-yolov7.md @@ -25,7 +25,7 @@ RTDETRv2 (Real-Time Detection Transformer version 2) builds upon the foundation ### Architectural Highlights -RTDETRv2 utilizes a hybrid encoder and a [transformer decoder](https://docs.ultralytics.com/reference/nn/modules/transformer/) architecture. By leveraging self-attention mechanisms, the model processes the entire image holistically, allowing it to understand complex spatial relationships better than strictly localized convolutional kernels. One of its most defining features is its natively NMS-free design. By eliminating Non-Maximum Suppression (NMS), RTDETRv2 removes a common bottleneck that introduces variable [inference latency](https://www.ultralytics.com/glossary/inference-latency) during deployment. +RTDETRv2 utilizes a hybrid encoder and a [transformer decoder](https://docs.ultralytics.com/reference/nn/modules/transformer) architecture. By leveraging self-attention mechanisms, the model processes the entire image holistically, allowing it to understand complex spatial relationships better than strictly localized convolutional kernels. One of its most defining features is its natively NMS-free design. By eliminating Non-Maximum Suppression (NMS), RTDETRv2 removes a common bottleneck that introduces variable [inference latency](https://www.ultralytics.com/glossary/inference-latency) during deployment. ### Strengths and Limitations @@ -51,11 +51,11 @@ YOLOv7's architecture is built around the concept of Extended Efficient Layer Ag ### Strengths and Limitations -YOLOv7 remains a highly capable model for standard [object detection](https://docs.ultralytics.com/tasks/detect/) tasks, offering excellent processing speeds on consumer GPUs. Its CNN nature means it typically requires less CUDA memory during training compared to transformer-based models like RTDETRv2. +YOLOv7 remains a highly capable model for standard [object detection](https://docs.ultralytics.com/tasks/detect) tasks, offering excellent processing speeds on consumer GPUs. Its CNN nature means it typically requires less CUDA memory during training compared to transformer-based models like RTDETRv2. -Despite these advantages, YOLOv7 still relies on NMS for post-processing. In environments with a high density of predictions, the NMS step can cause fluctuations in processing time, making strict real-time guarantees difficult. Additionally, compared to modern frameworks, the process of handling varied tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [pose estimation](https://docs.ultralytics.com/tasks/pose/) can be fragmented. +Despite these advantages, YOLOv7 still relies on NMS for post-processing. In environments with a high density of predictions, the NMS step can cause fluctuations in processing time, making strict real-time guarantees difficult. Additionally, compared to modern frameworks, the process of handling varied tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [pose estimation](https://docs.ultralytics.com/tasks/pose) can be fragmented. -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ## Performance Comparison @@ -79,16 +79,16 @@ Evaluating these models requires looking at the delicate balance between mean Av While RTDETRv2 and YOLOv7 were pivotal in pushing the boundaries of [computer vision applications](https://www.ultralytics.com/blog/60-impactful-computer-vision-applications), the AI landscape evolves rapidly. Released in January 2026, **[YOLO26](https://platform.ultralytics.com/ultralytics/yolo26)** synthesizes the best aspects of both CNN efficiency and transformer-like NMS-free architectures. -For developers and researchers building new systems, the integrated [Ultralytics Platform](https://docs.ultralytics.com/platform/) and Python ecosystem provide a unified experience that significantly reduces technical debt. +For developers and researchers building new systems, the integrated [Ultralytics Platform](https://docs.ultralytics.com/platform) and Python ecosystem provide a unified experience that significantly reduces technical debt. ### Key Innovations in YOLO26 -- **End-to-End NMS-Free Design:** YOLO26 is natively end-to-end, eliminating NMS post-processing for faster, simpler deployment. This breakthrough approach was first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), ensuring stable latency regardless of object density. +- **End-to-End NMS-Free Design:** YOLO26 is natively end-to-end, eliminating NMS post-processing for faster, simpler deployment. This breakthrough approach was first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), ensuring stable latency regardless of object density. - **Up to 43% Faster CPU Inference:** Specifically optimized for [edge computing](https://www.ultralytics.com/glossary/edge-computing) and devices without GPUs, making it far more versatile for field deployments than heavy transformer models. - **MuSGD Optimizer:** A hybrid of SGD and Muon (inspired by Moonshot AI's Kimi K2), bringing LLM training innovations to computer vision for more stable training and faster convergence. -- **DFL Removal:** Distribution Focal Loss has been removed, resulting in a simplified computational graph for smoother export to embedded NPUs and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) environments. +- **DFL Removal:** Distribution Focal Loss has been removed, resulting in a simplified computational graph for smoother export to embedded NPUs and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) environments. - **ProgLoss + STAL:** Improved loss functions yield notable enhancements in small-object recognition, which is critical for [robotics](https://www.ultralytics.com/glossary/robotics), IoT, and aerial imagery analysis. -- **Task-Specific Improvements:** YOLO26 isn't just for detection. It features multi-scale prototypes for segmentation, Residual Log-Likelihood Estimation (RLE) for pose tracking, and specialized angle loss addressing [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb/) boundary issues. +- **Task-Specific Improvements:** YOLO26 isn't just for detection. It features multi-scale prototypes for segmentation, Residual Log-Likelihood Estimation (RLE) for pose tracking, and specialized angle loss addressing [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb) boundary issues. ### Streamlined Developer Experience @@ -115,7 +115,7 @@ Selecting between these architectures depends heavily on the target hardware and ### When to Consider RTDETRv2 -RTDETRv2 is highly effective in [server-side processing](https://docs.ultralytics.com/guides/triton-inference-server/) environments equipped with powerful GPUs. Its global attention mechanism makes it suitable for complex scene understanding, such as highly crowded event monitoring or specialized medical imaging where overlapping features require deep contextual analysis. +RTDETRv2 is highly effective in [server-side processing](https://docs.ultralytics.com/guides/triton-inference-server) environments equipped with powerful GPUs. Its global attention mechanism makes it suitable for complex scene understanding, such as highly crowded event monitoring or specialized medical imaging where overlapping features require deep contextual analysis. ### When to Consider YOLOv7 @@ -123,8 +123,8 @@ YOLOv7 is often maintained in legacy academic research as a baseline comparison ### Why YOLO26 is the Recommended Standard -For modern [smart city](https://www.ultralytics.com/blog/computer-vision-ai-in-smart-cities) infrastructure, [drone navigation](https://www.ultralytics.com/blog/build-ai-powered-drone-applications-with-ultralytics-yolo11), and high-speed manufacturing, YOLO26 offers an unmatched balance. Its lower memory requirements make [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/) and training accessible on consumer hardware, while its NMS-free inference ensures rapid execution on constrained edge devices like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or NVIDIA Jetson. +For modern [smart city](https://www.ultralytics.com/blog/computer-vision-ai-in-smart-cities) infrastructure, [drone navigation](https://www.ultralytics.com/blog/build-ai-powered-drone-applications-with-ultralytics-yolo11), and high-speed manufacturing, YOLO26 offers an unmatched balance. Its lower memory requirements make [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning) and training accessible on consumer hardware, while its NMS-free inference ensures rapid execution on constrained edge devices like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or NVIDIA Jetson. !!! tip "Explore More Comparisons" - Interested in how these models stack up against other architectures? Check out our detailed guides on [YOLO11 vs. RTDETR](https://docs.ultralytics.com/compare/rtdetr-vs-yolo11/) and [YOLOv8 vs. YOLOv7](https://docs.ultralytics.com/compare/yolov8-vs-yolov7/) to find the perfect fit for your vision AI project. + Interested in how these models stack up against other architectures? Check out our detailed guides on [YOLO11 vs. RTDETR](https://docs.ultralytics.com/compare/rtdetr-vs-yolo11) and [YOLOv8 vs. YOLOv7](https://docs.ultralytics.com/compare/yolov8-vs-yolov7) to find the perfect fit for your vision AI project. diff --git a/docs/en/compare/rtdetr-vs-yolov8.md b/docs/en/compare/rtdetr-vs-yolov8.md index 42c1df49073..495be4239a8 100644 --- a/docs/en/compare/rtdetr-vs-yolov8.md +++ b/docs/en/compare/rtdetr-vs-yolov8.md @@ -32,9 +32,9 @@ At its core, RTDETRv2 leverages a hybrid architecture combining a CNN backbone w ### Weaknesses -Despite its powerful contextual understanding, transformer-based architectures like RTDETRv2 require immense computational overhead during training. They demand a significant amount of CUDA memory, making them difficult to train on consumer-grade hardware. Additionally, setting up a custom dataset and tuning the training hyperparameters often requires deep domain expertise, as the model lacks a highly polished, beginner-friendly software wrapper. Deployment to low-power edge devices such as older [Raspberry Pi hardware](https://docs.ultralytics.com/guides/raspberry-pi/) can also prove challenging due to the heavy attention mechanisms. +Despite its powerful contextual understanding, transformer-based architectures like RTDETRv2 require immense computational overhead during training. They demand a significant amount of CUDA memory, making them difficult to train on consumer-grade hardware. Additionally, setting up a custom dataset and tuning the training hyperparameters often requires deep domain expertise, as the model lacks a highly polished, beginner-friendly software wrapper. Deployment to low-power edge devices such as older [Raspberry Pi hardware](https://docs.ultralytics.com/guides/raspberry-pi) can also prove challenging due to the heavy attention mechanisms. -[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr){ .md-button } --- @@ -47,11 +47,11 @@ Since its release, [Ultralytics YOLOv8](https://platform.ultralytics.com/ultraly - **Authors:** Glenn Jocher, Ayush Chaurasia, and Jing Qiu - **Organization:** [Ultralytics](https://www.ultralytics.com/about) - **Date:** January 10, 2023 -- **Links:** [Official Documentation](https://docs.ultralytics.com/models/yolov8/) | [GitHub Repository](https://github.com/ultralytics/ultralytics) +- **Links:** [Official Documentation](https://docs.ultralytics.com/models/yolov8) | [GitHub Repository](https://github.com/ultralytics/ultralytics) ### Architecture and Strengths -YOLOv8 utilizes a highly optimized anchor-free CNN architecture with a decoupled head, significantly improving object localization and classification accuracy over previous generations. Its greatest strength lies in its incredible efficiency and versatility. The architecture requires substantially lower memory during training compared to vision transformers, allowing practitioners to run larger [batch sizes](https://www.ultralytics.com/glossary/batch-size) on standard GPUs. Furthermore, the Ultralytics ecosystem provides an unmatched, seamless workflow. The unified Python API enables [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/), training, validation, and export with just a few lines of code. +YOLOv8 utilizes a highly optimized anchor-free CNN architecture with a decoupled head, significantly improving object localization and classification accuracy over previous generations. Its greatest strength lies in its incredible efficiency and versatility. The architecture requires substantially lower memory during training compared to vision transformers, allowing practitioners to run larger [batch sizes](https://www.ultralytics.com/glossary/batch-size) on standard GPUs. Furthermore, the Ultralytics ecosystem provides an unmatched, seamless workflow. The unified Python API enables [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning), training, validation, and export with just a few lines of code. ### Weaknesses @@ -84,13 +84,13 @@ When comparing raw numbers, it becomes evident that both models prioritize diffe ### Memory Requirements and Training Efficiency -One of the most critical factors for independent developers and enterprise teams alike is training cost. Ultralytics YOLO models require significantly lower CUDA memory during the [training process](https://docs.ultralytics.com/modes/train/) than transformer architectures. A standard RTDETRv2 model may easily bottleneck a consumer GPU, whereas YOLOv8 converges quickly and reliably on hardware like the NVIDIA RTX 4070. +One of the most critical factors for independent developers and enterprise teams alike is training cost. Ultralytics YOLO models require significantly lower CUDA memory during the [training process](https://docs.ultralytics.com/modes/train) than transformer architectures. A standard RTDETRv2 model may easily bottleneck a consumer GPU, whereas YOLOv8 converges quickly and reliably on hardware like the NVIDIA RTX 4070. ## Ecosystem, API, and Ease of Use The true differentiator for modern AI solutions is the supporting software framework. The Ultralytics ecosystem simplifies complex engineering hurdles. With active development and robust community support on platforms like [Discord](https://discord.com/invite/ultralytics), YOLOv8 ensures your project doesn't stall due to poor documentation. -Furthermore, YOLOv8 goes beyond standard object detection. It is a true multi-task network with native support for [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). RTDETRv2 remains heavily focused purely on detection. +Furthermore, YOLOv8 goes beyond standard object detection. It is a true multi-task network with native support for [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Image Classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). RTDETRv2 remains heavily focused purely on detection. ### Code Example: Unified Simplicity @@ -111,7 +111,7 @@ results_cnn = model_cnn("https://ultralytics.com/images/bus.jpg") model_cnn.export(format="onnx") ``` -Once trained, YOLOv8 supports one-click exports to [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), and [OpenVINO](https://docs.ultralytics.com/integrations/openvino/), guaranteeing high-throughput inference across diverse hardware backends. +Once trained, YOLOv8 supports one-click exports to [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), and [OpenVINO](https://docs.ultralytics.com/integrations/openvino), guaranteeing high-throughput inference across diverse hardware backends. ## Use Cases and Recommendations @@ -129,17 +129,17 @@ RT-DETR is a strong choice for: YOLOv8 is recommended for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Looking Forward: The YOLO26 Advantage @@ -147,4 +147,4 @@ While YOLOv8 remains a legendary milestone, computer vision moves incredibly fas If you are drawn to the NMS-free design of RTDETRv2, YOLO26 incorporates a native **End-to-End NMS-Free Design**, combining the post-processing simplicity of transformers with the blazing speed of CNNs. Additionally, YOLO26 utilizes the groundbreaking **MuSGD Optimizer**, bringing LLM-style training stability to vision models for incredibly fast convergence. With **DFL Removal** (Distribution Focal Loss removed for simplified export and better edge/low-power device compatibility), YOLO26 achieves **up to 43% faster CPU inference**. Combined with advanced **ProgLoss + STAL** mechanisms for superior small-object detection, YOLO26 is definitively the recommended upgrade path over both YOLOv8 and RTDETRv2. -For further reading on alternative models, explore our guides on [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or read the detailed breakdown of [YOLOv10 vs YOLOv8](https://docs.ultralytics.com/compare/yolov10-vs-yolov8/) to see how NMS-free architecture evolved in the YOLO family. +For further reading on alternative models, explore our guides on [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or read the detailed breakdown of [YOLOv10 vs YOLOv8](https://docs.ultralytics.com/compare/yolov10-vs-yolov8) to see how NMS-free architecture evolved in the YOLO family. diff --git a/docs/en/compare/rtdetr-vs-yolov9.md b/docs/en/compare/rtdetr-vs-yolov9.md index 8128fef5673..80cd442bf69 100644 --- a/docs/en/compare/rtdetr-vs-yolov9.md +++ b/docs/en/compare/rtdetr-vs-yolov9.md @@ -26,7 +26,7 @@ Developed by researchers at Baidu, RTDETRv2 builds upon the original RT-DETR by A defining characteristic of RTDETRv2 is its natively **end-to-end NMS-free design**. By completely removing Non-Maximum Suppression (NMS) during post-processing, the model stabilizes inference latency and simplifies the deployment pipeline. The global attention mechanism allows the model to excel in complex scene understanding and dense crowds, as it evaluates the entire image context simultaneously. -[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr){ .md-button } ### YOLOv9: Programmable Gradient Information @@ -39,7 +39,7 @@ YOLOv9, a highly efficient CNN-based architecture, tackles the information bottl YOLOv9 relies on the proven [convolutional neural network](https://en.wikipedia.org/wiki/Convolutional_neural_network) foundations but maximizes parameter efficiency. By retaining crucial information during the feed-forward process, it ensures reliable weight updates, resulting in an incredibly lightweight yet highly accurate model. However, unlike RTDETRv2, YOLOv9 still relies on standard NMS post-processing. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ## Performance and Resource Efficiency @@ -68,11 +68,11 @@ Transformers like RTDETRv2 are notoriously memory-intensive during training, oft ## The Ultralytics Advantage: Ecosystem and Ease of Use -While researching standalone repositories like the official RTDETRv2 or YOLOv9 GitHub pages can be highly educational, production environments demand stability, ease of use, and a well-maintained ecosystem. Integrating these models through the [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) offers a seamless developer experience. +While researching standalone repositories like the official RTDETRv2 or YOLOv9 GitHub pages can be highly educational, production environments demand stability, ease of use, and a well-maintained ecosystem. Integrating these models through the [Ultralytics Python API](https://docs.ultralytics.com/usage/python) offers a seamless developer experience. ### Unified API and Versatility -The Ultralytics framework abstracts away the complexities of data loading, augmentations, and distributed training. Furthermore, while the original RTDETRv2 is strictly focused on detection, the Ultralytics ecosystem allows users to easily transition between [Object Detection](https://docs.ultralytics.com/tasks/detect/), [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), and [Pose Estimation](https://docs.ultralytics.com/tasks/pose/). +The Ultralytics framework abstracts away the complexities of data loading, augmentations, and distributed training. Furthermore, while the original RTDETRv2 is strictly focused on detection, the Ultralytics ecosystem allows users to easily transition between [Object Detection](https://docs.ultralytics.com/tasks/detect), [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), and [Pose Estimation](https://docs.ultralytics.com/tasks/pose). ```python from ultralytics import RTDETR, YOLO @@ -89,7 +89,7 @@ results = model_rtdetr.predict("https://ultralytics.com/images/bus.jpg") model_yolo.export(format="engine") ``` -With robust documentation, automatic [experiment tracking](https://docs.ultralytics.com/integrations/comet/), and seamless [export capabilities](https://docs.ultralytics.com/modes/export/) to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), and OpenVINO, Ultralytics drastically reduces the time from prototype to production. +With robust documentation, automatic [experiment tracking](https://docs.ultralytics.com/integrations/comet), and seamless [export capabilities](https://docs.ultralytics.com/modes/export) to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), and OpenVINO, Ultralytics drastically reduces the time from prototype to production. ## Ideal Use Cases @@ -106,7 +106,7 @@ Thanks to its global attention mechanism, RTDETRv2 is a powerhouse for **server- YOLOv9 is a champion of **resource-constrained edge deployments**. Its computational efficiency makes it ideal for: - **Robotics:** Real-time navigation and obstacle avoidance where minimal latency is required. -- **Smart City IoT:** Deploying on edge devices like the [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) for traffic monitoring. +- **Smart City IoT:** Deploying on edge devices like the [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) for traffic monitoring. - **Industrial Inspection:** High-speed assembly line quality control requiring high frames-per-second (FPS). ## The Future: Enter Ultralytics YOLO26 diff --git a/docs/en/compare/rtdetr-vs-yolox.md b/docs/en/compare/rtdetr-vs-yolox.md index febdbdd64b4..7c3a5089634 100644 --- a/docs/en/compare/rtdetr-vs-yolox.md +++ b/docs/en/compare/rtdetr-vs-yolox.md @@ -36,7 +36,7 @@ RTDETRv2 relies heavily on the self-attention mechanisms inherent to transformer The primary strength of RTDETRv2 lies in its native end-to-end design. By skipping NMS, it avoids the latency spikes often associated with dense overlapping predictions. However, the heavy computational footprint of its transformer blocks means that it demands substantial GPU resources for both training and deployment. This makes it less ideal for resource-constrained edge devices or legacy mobile hardware. -[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr){ .md-button } ## YOLOX: Advancing Anchor-Free CNNs @@ -83,14 +83,14 @@ While both RTDETRv2 and YOLOX offer unique benefits, modern developers often req ### Key Innovations of YOLO26 -- **End-to-End NMS-Free Design:** Building on concepts first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 operates natively without NMS. This delivers the seamless inference of RTDETRv2 without the crushing memory requirements of transformers. +- **End-to-End NMS-Free Design:** Building on concepts first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 operates natively without NMS. This delivers the seamless inference of RTDETRv2 without the crushing memory requirements of transformers. - **MuSGD Optimizer:** Inspired by large language model training innovations, the hybrid MuSGD optimizer (blending SGD and Muon) stabilizes the training process and drastically accelerates convergence. - **Up to 43% Faster CPU Inference:** By strategically removing the Distribution Focal Loss (DFL) module, YOLO26 is specifically optimized for edge computing and low-power devices, making it substantially faster on CPUs than previous iterations like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11). - **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, addressing a common pain point in aerial imagery and [robotics applications](https://www.ultralytics.com/solutions/ai-in-robotics). ### Unmatched Versatility and Ecosystem -Beyond raw performance, the [Ultralytics Platform](https://platform.ultralytics.com) offers a comprehensive, zero-to-production ecosystem. Unlike static academic repositories, Ultralytics models are actively maintained and seamlessly support multiple tasks from a single, intuitive API. Whether you are performing [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), tracking poses via [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), or handling rotated objects with [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), the workflow remains identical. +Beyond raw performance, the [Ultralytics Platform](https://platform.ultralytics.com) offers a comprehensive, zero-to-production ecosystem. Unlike static academic repositories, Ultralytics models are actively maintained and seamlessly support multiple tasks from a single, intuitive API. Whether you are performing [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), tracking poses via [Pose Estimation](https://docs.ultralytics.com/tasks/pose), or handling rotated objects with [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb), the workflow remains identical. Furthermore, Ultralytics models are renowned for their low memory requirements during both training and inference, allowing researchers to run larger batch sizes on consumer-grade hardware—a stark contrast to the heavy footprint of transformer-based architectures. diff --git a/docs/en/compare/yolo11-vs-damo-yolo.md b/docs/en/compare/yolo11-vs-damo-yolo.md index 19137cc86b2..4f586381707 100644 --- a/docs/en/compare/yolo11-vs-damo-yolo.md +++ b/docs/en/compare/yolo11-vs-damo-yolo.md @@ -23,11 +23,11 @@ Developed by the team at Ultralytics, **YOLO11** represents a highly refined ite - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2024-09-27 - **GitHub:** [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11/) +- **Docs:** [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11) [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } -YOLO11 shines in its versatility. While many traditional models focus solely on bounding boxes, YOLO11 natively supports [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/). This multi-modal capability allows developers to consolidate their [vision AI](https://www.ultralytics.com/glossary/computer-vision-cv) pipelines under a single, well-maintained framework. +YOLO11 shines in its versatility. While many traditional models focus solely on bounding boxes, YOLO11 natively supports [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose). This multi-modal capability allows developers to consolidate their [vision AI](https://www.ultralytics.com/glossary/computer-vision-cv) pipelines under a single, well-maintained framework. ### DAMO-YOLO @@ -46,7 +46,7 @@ The core philosophy of DAMO-YOLO revolves around rep-parameterization and automa !!! note "Other Models to Consider" - While comparing YOLO11 and DAMO-YOLO, consider checking out the newer [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26). It introduces natively end-to-end NMS-free inference and delivers up to 43% faster CPU speeds. You might also explore comparisons involving [YOLOX](https://docs.ultralytics.com/compare/yolox-vs-damo-yolo/) or [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8). + While comparing YOLO11 and DAMO-YOLO, consider checking out the newer [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26). It introduces natively end-to-end NMS-free inference and delivers up to 43% faster CPU speeds. You might also explore comparisons involving [YOLOX](https://docs.ultralytics.com/compare/yolox-vs-damo-yolo) or [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8). ## Performance and Architecture Comparison @@ -75,7 +75,7 @@ Conversely, **DAMO-YOLO**'s MAE-NAS generated backbones are finely tuned for hig When factoring in development time, the **Ease of Use** of a model becomes just as important as its raw benchmarks. -**YOLO11** is built heavily on the principle of developer accessibility. The comprehensive `ultralytics` package abstracts away the heavy lifting of dataset parsing, augmentation, and hyperparameter tuning. Exporting models to production formats like [ONNX](https://onnx.ai/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), and [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) requires only a single command. +**YOLO11** is built heavily on the principle of developer accessibility. The comprehensive `ultralytics` package abstracts away the heavy lifting of dataset parsing, augmentation, and hyperparameter tuning. Exporting models to production formats like [ONNX](https://onnx.ai/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), and [OpenVINO](https://docs.ultralytics.com/integrations/openvino) requires only a single command. ```python from ultralytics import YOLO @@ -100,9 +100,9 @@ Choosing between YOLO11 and DAMO-YOLO depends on your specific project requireme YOLO11 is a strong choice for: -- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) where reliability and active maintenance are paramount. -- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [OBB](https://docs.ultralytics.com/tasks/obb/) within a single unified framework. -- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python/). +- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) where reliability and active maintenance are paramount. +- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [OBB](https://docs.ultralytics.com/tasks/obb) within a single unified framework. +- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python). ### When to Choose DAMO-YOLO @@ -114,21 +114,21 @@ DAMO-YOLO is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Real-World Applications and Use Cases ### Autonomous Systems and Drones -For aerial imagery and UAV deployments, **YOLO11** provides an incredibly favorable performance balance. Small object detection is a massive hurdle in drone analytics, but YOLO11 handles varying scales natively out of the box. Additionally, the low [memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics/) allow YOLO11 Nano and Small variants to run directly on lightweight edge CPUs or NPUs strapped to the drone. +For aerial imagery and UAV deployments, **YOLO11** provides an incredibly favorable performance balance. Small object detection is a massive hurdle in drone analytics, but YOLO11 handles varying scales natively out of the box. Additionally, the low [memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics) allow YOLO11 Nano and Small variants to run directly on lightweight edge CPUs or NPUs strapped to the drone. ### Industrial Automation and Quality Control -In smart factories, latency is paramount. While **DAMO-YOLO** offers robust inference speeds on heavy server-grade GPUs due to its RepGFPN neck, the rigid integration can be overkill. YOLO11 often acts as a superior alternative for automated quality control due to its simple [tracking APIs](https://docs.ultralytics.com/modes/track/) and the ability to seamlessly pivot from pure detection to [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb/) tasks if the defects require angled boundary recognition. +In smart factories, latency is paramount. While **DAMO-YOLO** offers robust inference speeds on heavy server-grade GPUs due to its RepGFPN neck, the rigid integration can be overkill. YOLO11 often acts as a superior alternative for automated quality control due to its simple [tracking APIs](https://docs.ultralytics.com/modes/track) and the ability to seamlessly pivot from pure detection to [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb) tasks if the defects require angled boundary recognition. ### Smart Healthcare and Medical Imaging diff --git a/docs/en/compare/yolo11-vs-efficientdet.md b/docs/en/compare/yolo11-vs-efficientdet.md index c9965985c1f..f1ed8cc6f38 100644 --- a/docs/en/compare/yolo11-vs-efficientdet.md +++ b/docs/en/compare/yolo11-vs-efficientdet.md @@ -6,7 +6,7 @@ keywords: YOLO11, EfficientDet, object detection, model comparison, YOLO vs Effi # YOLO11 vs EfficientDet: A Comprehensive Technical Comparison -Selecting the optimal neural network for [computer vision](https://en.wikipedia.org/wiki/Computer_vision) projects requires a deep understanding of the available architectures. This guide provides an in-depth technical comparison between [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and Google's EfficientDet. We will explore their architectural differences, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/), training efficiencies, and ideal deployment scenarios to help you make an informed decision for your [machine learning](https://en.wikipedia.org/wiki/Machine_learning) workloads. +Selecting the optimal neural network for [computer vision](https://en.wikipedia.org/wiki/Computer_vision) projects requires a deep understanding of the available architectures. This guide provides an in-depth technical comparison between [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and Google's EfficientDet. We will explore their architectural differences, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics), training efficiencies, and ideal deployment scenarios to help you make an informed decision for your [machine learning](https://en.wikipedia.org/wiki/Machine_learning) workloads. @@ -23,7 +23,7 @@ Authors: Glenn Jocher and Jing Qiu Organization: [Ultralytics](https://www.ultralytics.com/) Date: 2024-09-27 GitHub: [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -Docs: [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11/) +Docs: [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11) [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } @@ -40,7 +40,7 @@ Docs: [https://github.com/google/automl/tree/master/efficientdet#readme](https:/ !!! tip "Ecosystem Advantage" - When working with computer vision models, the surrounding ecosystem is just as important as the model itself. The [Ultralytics ecosystem](https://docs.ultralytics.com/integrations/) provides an unparalleled developer experience, offering extensive documentation, active community support, and seamless export capabilities to formats like [ONNX](https://onnx.ai/) and [TensorRT](https://developer.nvidia.com/tensorrt). + When working with computer vision models, the surrounding ecosystem is just as important as the model itself. The [Ultralytics ecosystem](https://docs.ultralytics.com/integrations) provides an unparalleled developer experience, offering extensive documentation, active community support, and seamless export capabilities to formats like [ONNX](https://onnx.ai/) and [TensorRT](https://developer.nvidia.com/tensorrt). ## Architectural Innovations @@ -54,7 +54,7 @@ While highly efficient for its time, EfficientDet's reliance on the TensorFlow [ YOLO11 represents a significant leap forward in [object detection architectures](https://www.ultralytics.com/glossary/object-detection-architectures). It builds upon the successes of its predecessors, introducing refined C3k2 blocks and an improved [Spatial Pyramid Pooling](https://arxiv.org/abs/1406.4729) module. These enhancements lead to superior [feature extraction](https://www.ultralytics.com/glossary/feature-extraction), allowing YOLO11 to capture intricate visual patterns with exceptional clarity. -A major advantage of YOLO11 is its **versatility**. While EfficientDet is strictly an [object detection](https://docs.ultralytics.com/tasks/detect/) model, YOLO11 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). Furthermore, YOLO11 boasts incredibly low **memory requirements** during both training and inference, making it vastly superior to older models and bulky [vision transformers](https://arxiv.org/abs/2010.11929) when deploying to resource-constrained [edge AI](https://www.ultralytics.com/glossary/edge-ai) environments. +A major advantage of YOLO11 is its **versatility**. While EfficientDet is strictly an [object detection](https://docs.ultralytics.com/tasks/detect) model, YOLO11 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb). Furthermore, YOLO11 boasts incredibly low **memory requirements** during both training and inference, making it vastly superior to older models and bulky [vision transformers](https://arxiv.org/abs/2010.11929) when deploying to resource-constrained [edge AI](https://www.ultralytics.com/glossary/edge-ai) environments. ## Performance and Benchmarks @@ -83,7 +83,7 @@ As shown, YOLO11 achieves a highly favorable **performance balance**. YOLO11x ac One of the defining characteristics of Ultralytics models is their **ease of use**. Training an EfficientDet model often requires navigating complex TensorFlow graph configurations and managing intricate dependency chains. In stark contrast, YOLO11 is built on a clean, thoroughly modern [PyTorch](https://pytorch.org/) foundation. -This **well-maintained ecosystem** means developers can install the package, load a pre-trained model, and start training on a custom [dataset](https://docs.ultralytics.com/datasets/) in just a few lines of code. +This **well-maintained ecosystem** means developers can install the package, load a pre-trained model, and start training on a custom [dataset](https://docs.ultralytics.com/datasets) in just a few lines of code. ### Python Code Example @@ -120,7 +120,7 @@ Key YOLO26 innovations include: !!! tip "Alternative Models to Explore" - If your project has highly specific requirements, you might also want to benchmark the [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) model for transformer-based detection, or the widely adopted [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), which remains a staple in many legacy enterprise deployments. + If your project has highly specific requirements, you might also want to benchmark the [RT-DETR](https://docs.ultralytics.com/models/rtdetr) model for transformer-based detection, or the widely adopted [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), which remains a staple in many legacy enterprise deployments. ## Use Cases and Recommendations @@ -130,9 +130,9 @@ Choosing between YOLO11 and EfficientDet depends on your specific project requir YOLO11 is a strong choice for: -- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) where reliability and active maintenance are paramount. -- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [OBB](https://docs.ultralytics.com/tasks/obb/) within a single unified framework. -- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python/). +- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) where reliability and active maintenance are paramount. +- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [OBB](https://docs.ultralytics.com/tasks/obb) within a single unified framework. +- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python). ### When to Choose EfficientDet @@ -144,14 +144,14 @@ EfficientDet is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Conclusion EfficientDet was a pioneering architecture that proved the viability of compound scaling in object detection. However, the rapid pace of AI research has brought forth models that are simply more capable, easier to integrate, and faster to run. -With its robust multi-task capabilities, incredible GPU inference speeds, and arguably the most developer-friendly API in the industry, **YOLO11** is the clear winner for modern vision pipelines. For those aiming at the absolute bleeding edge of technology—especially for edge-first deployments—upgrading to [YOLO26](https://docs.ultralytics.com/models/yolo26/) provides the ultimate combination of NMS-free speed and unparalleled accuracy. +With its robust multi-task capabilities, incredible GPU inference speeds, and arguably the most developer-friendly API in the industry, **YOLO11** is the clear winner for modern vision pipelines. For those aiming at the absolute bleeding edge of technology—especially for edge-first deployments—upgrading to [YOLO26](https://docs.ultralytics.com/models/yolo26) provides the ultimate combination of NMS-free speed and unparalleled accuracy. diff --git a/docs/en/compare/yolo11-vs-pp-yoloe.md b/docs/en/compare/yolo11-vs-pp-yoloe.md index 332dd0f015c..dc117cc2a4f 100644 --- a/docs/en/compare/yolo11-vs-pp-yoloe.md +++ b/docs/en/compare/yolo11-vs-pp-yoloe.md @@ -6,7 +6,7 @@ keywords: YOLO11, PP-YOLOE+, object detection, YOLO comparison, real-time detect # YOLO11 vs PP-YOLOE+: A Technical Comparison of Real-Time Detectors -Selecting the optimal neural network architecture is critical when deploying [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) applications in production. In this technical comparison, we examine two prominent models in the real-time object detection space: [Ultralytics YOLO11](https://docs.ultralytics.com/models/yolo11/) and Baidu's PP-YOLOE+. Both architectures offer robust performance, but they approach the challenges of accuracy, inference speed, and developer ecosystem quite differently. +Selecting the optimal neural network architecture is critical when deploying [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) applications in production. In this technical comparison, we examine two prominent models in the real-time object detection space: [Ultralytics YOLO11](https://docs.ultralytics.com/models/yolo11) and Baidu's PP-YOLOE+. Both architectures offer robust performance, but they approach the challenges of accuracy, inference speed, and developer ecosystem quite differently. Below is an interactive chart showcasing the performance boundaries of these models to help you identify the best fit for your hardware constraints. @@ -27,7 +27,7 @@ Developed by Ultralytics, YOLO11 represents a highly refined iteration of the YO - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2024-09-27 - **GitHub:** [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11/) +- **Docs:** [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11) [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } @@ -48,7 +48,7 @@ PP-YOLOE+ is an evolved version of PP-YOLOv2, built upon the PaddlePaddle framew The fundamental architectural designs of YOLO11 and PP-YOLOE+ reflect their differing priorities in the [computer vision](https://www.ultralytics.com/blog/all-you-need-to-know-about-computer-vision-tasks) landscape. -**YOLO11** builds upon a highly optimized backbone and an anchor-free detection head. It utilizes C3k2 blocks and Spatial Pyramid Pooling - Fast (SPPF) to capture multi-scale features with minimal computational overhead. This design is highly advantageous for reducing [inference latency](https://www.ultralytics.com/glossary/inference-latency) on resource-constrained devices like edge NPUs and mobile CPUs. Furthermore, YOLO11 is designed natively for multi-task learning, supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding box (OBB) detection](https://docs.ultralytics.com/tasks/obb/) right out of the box. +**YOLO11** builds upon a highly optimized backbone and an anchor-free detection head. It utilizes C3k2 blocks and Spatial Pyramid Pooling - Fast (SPPF) to capture multi-scale features with minimal computational overhead. This design is highly advantageous for reducing [inference latency](https://www.ultralytics.com/glossary/inference-latency) on resource-constrained devices like edge NPUs and mobile CPUs. Furthermore, YOLO11 is designed natively for multi-task learning, supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding box (OBB) detection](https://docs.ultralytics.com/tasks/obb) right out of the box. **PP-YOLOE+** introduces the CSPRepResNet backbone and an Efficient Task-aligned head (ET-head). It heavily utilizes rep-parameterization techniques to increase representational capacity during training while folding those parameters into standard convolutions for inference. While this yields impressive [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map), the resulting models tend to be heavier in terms of parameters and memory footprint, making them better suited for deployment on robust server GPUs rather than lightweight edge devices. @@ -76,15 +76,15 @@ When evaluating performance, we look at accuracy (mAP), inference speed across d ### Analysis -YOLO11 demonstrates a clear advantage in **performance balance** and parameter efficiency. For instance, `YOLO11m` achieves a higher mAP (51.5) than `PP-YOLOE+m` (49.8) while utilizing fewer parameters (20.1M vs 23.43M) and achieving significantly faster inference speeds on TensorRT (4.7ms vs 5.56ms). The lightweight nature of YOLO11 models inherently translates to lower memory requirements during both [model training](https://docs.ultralytics.com/modes/train/) and deployment. +YOLO11 demonstrates a clear advantage in **performance balance** and parameter efficiency. For instance, `YOLO11m` achieves a higher mAP (51.5) than `PP-YOLOE+m` (49.8) while utilizing fewer parameters (20.1M vs 23.43M) and achieving significantly faster inference speeds on TensorRT (4.7ms vs 5.56ms). The lightweight nature of YOLO11 models inherently translates to lower memory requirements during both [model training](https://docs.ultralytics.com/modes/train) and deployment. ## Training Ecosystem and Ease of Use -The true value of a model often lies in how easily developers can train it on custom [computer vision datasets](https://docs.ultralytics.com/datasets/) and deploy it to production. +The true value of a model often lies in how easily developers can train it on custom [computer vision datasets](https://docs.ultralytics.com/datasets) and deploy it to production. ### The Ultralytics Advantage -Ultralytics prioritizes a streamlined developer experience. Training YOLO11 is managed through a simple Python API or CLI, abstracting away complex boilerplate code. The [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolo11) further enhances this by providing no-code training, automated dataset management, and single-click exports to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/), CoreML, and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). +Ultralytics prioritizes a streamlined developer experience. Training YOLO11 is managed through a simple Python API or CLI, abstracting away complex boilerplate code. The [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolo11) further enhances this by providing no-code training, automated dataset management, and single-click exports to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx), CoreML, and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). Furthermore, YOLO models are highly memory-efficient during training, avoiding the massive VRAM overheads typical of transformer-based architectures or heavy rep-parameterized models, enabling training on consumer-grade hardware. @@ -119,12 +119,12 @@ While YOLO11 remains incredibly powerful, the field of AI moves fast. For the ab **Key YOLO26 Innovations:** -- **End-to-End NMS-Free Design:** YOLO26 natively eliminates [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing. This significantly speeds up inference and simplifies deployment logic, an architectural leap first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/). +- **End-to-End NMS-Free Design:** YOLO26 natively eliminates [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing. This significantly speeds up inference and simplifies deployment logic, an architectural leap first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10). - **Up to 43% Faster CPU Inference:** Optimized specifically for edge devices without GPUs, ensuring real-time performance on lower-power hardware. - **MuSGD Optimizer:** Inspired by LLM training stability, this hybrid of SGD and Muon ensures faster convergence and more stable training. -- **ProgLoss + STAL:** Improved loss functions drastically enhance small-object recognition, which is critical for [drone applications](https://docs.ultralytics.com/datasets/detect/visdrone/) and security surveillance. +- **ProgLoss + STAL:** Improved loss functions drastically enhance small-object recognition, which is critical for [drone applications](https://docs.ultralytics.com/datasets/detect/visdrone) and security surveillance. - **DFL Removal:** The removal of Distribution Focal Loss simplifies model export and dramatically improves compatibility across a wide range of edge devices. For new projects prioritizing speed, seamless export, and maximum accuracy, we highly recommend leveraging the capabilities of YOLO26 via the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolo26). -If you are evaluating other architectures, you may also be interested in comparing YOLO11 to [RT-DETR](https://docs.ultralytics.com/compare/rtdetr-vs-yolo11/) or exploring how the legacy [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) holds up in modern benchmarks. +If you are evaluating other architectures, you may also be interested in comparing YOLO11 to [RT-DETR](https://docs.ultralytics.com/compare/rtdetr-vs-yolo11) or exploring how the legacy [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) holds up in modern benchmarks. diff --git a/docs/en/compare/yolo11-vs-rtdetr.md b/docs/en/compare/yolo11-vs-rtdetr.md index 948f080ea18..e662e7d56d2 100644 --- a/docs/en/compare/yolo11-vs-rtdetr.md +++ b/docs/en/compare/yolo11-vs-rtdetr.md @@ -17,13 +17,13 @@ By analyzing their architectures, performance metrics, and ideal deployment scen ## YOLO11: The Benchmark for Real-World Versatility -Introduced by Ultralytics, YOLO11 builds upon years of foundational research to deliver a model that is fast, accurate, and incredibly versatile. It is engineered to seamlessly handle [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb/) extraction natively. +Introduced by Ultralytics, YOLO11 builds upon years of foundational research to deliver a model that is fast, accurate, and incredibly versatile. It is engineered to seamlessly handle [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb) extraction natively. - **Authors:** Glenn Jocher and Jing Qiu - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2024-09-27 - **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11/) +- **Docs:** [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11) [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } @@ -31,7 +31,7 @@ Introduced by Ultralytics, YOLO11 builds upon years of foundational research to YOLO11 features a refined CNN backbone and advanced spatial feature pyramids, making it exceptionally resource-efficient. It thrives in environments with strict hardware constraints, offering a minimal memory footprint during both training and inference. The [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolo11) provides native support for YOLO11, enabling streamlined model monitoring, data annotation, and cloud training without needing to stitch together disparate MLops tools. -For developers targeting [edge computing](https://docs.ultralytics.com/guides/model-deployment-options/), YOLO11 boasts ultra-low latency. Its lightweight nature allows it to run efficiently on devices ranging from Raspberry Pis to consumer-grade mobile phones, making it a standard for smart retail, [manufacturing quality control](https://www.ultralytics.com/solutions/ai-in-manufacturing), and automated traffic management. +For developers targeting [edge computing](https://docs.ultralytics.com/guides/model-deployment-options), YOLO11 boasts ultra-low latency. Its lightweight nature allows it to run efficiently on devices ranging from Raspberry Pis to consumer-grade mobile phones, making it a standard for smart retail, [manufacturing quality control](https://www.ultralytics.com/solutions/ai-in-manufacturing), and automated traffic management. ## RTDETRv2: Real-Time Transformers by Baidu @@ -44,17 +44,17 @@ RTDETRv2 (Real-Time Detection Transformer version 2) represents Baidu's effort t - **GitHub:** [RT-DETRv2 Repository](https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetrv2_pytorch) - **Docs:** [RTDETRv2 README](https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetrv2_pytorch#readme) -[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr){ .md-button } ### Architecture and Strengths Unlike traditional CNNs, RTDETRv2 employs an encoder-decoder architecture with self-attention mechanisms, allowing it to capture global context across an image. This is particularly advantageous in crowded scenes where occlusions are frequent. RTDETRv2 eliminates the need for Non-Maximum Suppression (NMS) in post-processing, relying instead on Hungarian matching during training for one-to-one bipartite matching. -However, transformer models are notoriously hungry for [VRAM and CUDA memory](https://docs.ultralytics.com/guides/yolo-performance-metrics/). Training RTDETRv2 from scratch or fine-tuning on custom datasets often requires substantial high-end GPU clusters, which can be a barrier for smaller agile teams compared to the lightweight training footprint of Ultralytics models. +However, transformer models are notoriously hungry for [VRAM and CUDA memory](https://docs.ultralytics.com/guides/yolo-performance-metrics). Training RTDETRv2 from scratch or fine-tuning on custom datasets often requires substantial high-end GPU clusters, which can be a barrier for smaller agile teams compared to the lightweight training footprint of Ultralytics models. ## Performance and Metrics Analysis -When evaluating these models on the standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/), we observe clear trade-offs between parameters, FLOPs, and raw accuracy. +When evaluating these models on the standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco), we observe clear trade-offs between parameters, FLOPs, and raw accuracy. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -73,7 +73,7 @@ When evaluating these models on the standard [COCO dataset](https://docs.ultraly As seen in the table, YOLO11 provides an incredible performance-to-size ratio. The YOLO11x achieves a higher mAPval (54.7) compared to RTDETRv2-x (54.3), while using significantly fewer parameters (56.9M vs 76M) and vastly fewer computational FLOPs (194.9B vs 259B). -Furthermore, YOLO11's inference speeds on T4 [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) are exceptionally fast. YOLO11s completes inference in just 2.5ms, whereas the smallest RTDETRv2-s takes 5.03ms. This makes YOLO11 the definitive choice for high-speed, real-time video analytics streams where frame processing time is the primary bottleneck. +Furthermore, YOLO11's inference speeds on T4 [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) are exceptionally fast. YOLO11s completes inference in just 2.5ms, whereas the smallest RTDETRv2-s takes 5.03ms. This makes YOLO11 the definitive choice for high-speed, real-time video analytics streams where frame processing time is the primary bottleneck. !!! note "The Cost of Transformers" @@ -83,7 +83,7 @@ Furthermore, YOLO11's inference speeds on T4 [TensorRT](https://docs.ultralytics The core advantage of adopting an Ultralytics model lies in the surrounding ecosystem. Training RTDETRv2 often involves navigating complex research-grade repositories, adjusting intricate bipartite matching loss weights, and managing significant memory overhead. -Conversely, Ultralytics focuses heavily on developer experience. The unified Python API abstracts away boilerplate code, integrating seamlessly with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) for [experiment tracking](https://www.ultralytics.com/glossary/experiment-tracking), and handling data augmentations automatically. +Conversely, Ultralytics focuses heavily on developer experience. The unified Python API abstracts away boilerplate code, integrating seamlessly with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) for [experiment tracking](https://www.ultralytics.com/glossary/experiment-tracking), and handling data augmentations automatically. Here is how simple it is to train and export a model using the `ultralytics` package: @@ -105,11 +105,11 @@ train_results = model.train( export_path = model.export(format="onnx") ``` -Once trained, exporting a YOLO11 model to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/), [OpenVINO](https://docs.ultralytics.com/integrations/openvino/), or [CoreML](https://docs.ultralytics.com/integrations/coreml/) requires only a single command, ensuring your vision pipeline can scale effortlessly across diverse hardware backends. +Once trained, exporting a YOLO11 model to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx), [OpenVINO](https://docs.ultralytics.com/integrations/openvino), or [CoreML](https://docs.ultralytics.com/integrations/coreml) requires only a single command, ensuring your vision pipeline can scale effortlessly across diverse hardware backends. !!! tip "Multi-Task Capabilities" - Remember that while RTDETRv2 focuses exclusively on bounding box detection, the YOLO11 architecture natively supports [pose estimation](https://docs.ultralytics.com/tasks/pose/) and [instance segmentation](https://docs.ultralytics.com/tasks/segment/), allowing you to consolidate multiple vision tasks into a single model family. + Remember that while RTDETRv2 focuses exclusively on bounding box detection, the YOLO11 architecture natively supports [pose estimation](https://docs.ultralytics.com/tasks/pose) and [instance segmentation](https://docs.ultralytics.com/tasks/segment), allowing you to consolidate multiple vision tasks into a single model family. ## Use Cases and Recommendations @@ -119,9 +119,9 @@ Choosing between YOLO11 and RT-DETR depends on your specific project requirement YOLO11 is a strong choice for: -- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) where reliability and active maintenance are paramount. -- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [OBB](https://docs.ultralytics.com/tasks/obb/) within a single unified framework. -- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python/). +- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) where reliability and active maintenance are paramount. +- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [OBB](https://docs.ultralytics.com/tasks/obb) within a single unified framework. +- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python). ### When to Choose RT-DETR @@ -133,11 +133,11 @@ RT-DETR is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Looking Ahead: The Power of YOLO26 @@ -152,4 +152,4 @@ YOLO26 also introduces several revolutionary features: [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } -For those interested in exploring a wider range of architectures, the Ultralytics documentation also provides insights into [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), the widely adopted [YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5), and specialized models like [YOLO-World](https://docs.ultralytics.com/models/yolo-world/) for open-vocabulary detection applications. Ultimately, whether prioritizing the proven stability of YOLO11 or the breakthrough innovations of YOLO26, the Ultralytics ecosystem delivers unparalleled tools to bring your computer vision solutions to life. +For those interested in exploring a wider range of architectures, the Ultralytics documentation also provides insights into [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), the widely adopted [YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5), and specialized models like [YOLO-World](https://docs.ultralytics.com/models/yolo-world) for open-vocabulary detection applications. Ultimately, whether prioritizing the proven stability of YOLO11 or the breakthrough innovations of YOLO26, the Ultralytics ecosystem delivers unparalleled tools to bring your computer vision solutions to life. diff --git a/docs/en/compare/yolo11-vs-yolo26.md b/docs/en/compare/yolo11-vs-yolo26.md index d189a77ae47..38946559d36 100644 --- a/docs/en/compare/yolo11-vs-yolo26.md +++ b/docs/en/compare/yolo11-vs-yolo26.md @@ -17,7 +17,7 @@ Whether you are building [drone delivery systems](https://www.amazon.com/b?ie=UT ## Model Lineage and Ecosystem -Both models benefit from the comprehensive [Ultralytics ecosystem](https://github.com/ultralytics/ultralytics), characterized by its straightforward API, continuous maintenance, and a vibrant community. They offer unmatched versatility, naturally supporting [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) tasks out of the box. +Both models benefit from the comprehensive [Ultralytics ecosystem](https://github.com/ultralytics/ultralytics), characterized by its straightforward API, continuous maintenance, and a vibrant community. They offer unmatched versatility, naturally supporting [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) tasks out of the box. ### YOLO11: The Established Standard @@ -27,7 +27,7 @@ Released in late 2024, YOLO11 refined the advancements of earlier generations, c - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2024-09-27 - **GitHub:** [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11/) +- **Docs:** [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11) [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } @@ -39,7 +39,7 @@ Introduced in early 2026, YOLO26 represents a paradigm shift in edge computing a - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2026-01-14 - **GitHub:** [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26/) +- **Docs:** [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26) [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -53,7 +53,7 @@ While YOLO11 relies on traditional post-processing methods that have powered com ### End-to-End NMS-Free Design -One of the most significant upgrades in YOLO26 is its natively end-to-end architecture. It eliminates Non-Maximum Suppression (NMS) post-processing, a concept first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/). Bypassing NMS drastically simplifies the deployment pipeline and guarantees consistent latency, which is essential for real-time applications like [autonomous driving algorithms](https://waymo.com/research/). +One of the most significant upgrades in YOLO26 is its natively end-to-end architecture. It eliminates Non-Maximum Suppression (NMS) post-processing, a concept first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10). Bypassing NMS drastically simplifies the deployment pipeline and guarantees consistent latency, which is essential for real-time applications like [autonomous driving algorithms](https://waymo.com/research/). ### DFL Removal for Edge Optimization @@ -89,7 +89,7 @@ As demonstrated, the YOLO26 Nano (YOLO26n) jumps significantly in accuracy while !!! tip "Exporting for Maximum Speed" - To squeeze every drop of performance from these models, export them using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) on NVIDIA hardware or [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) for Intel CPUs. The NMS-free design of YOLO26 makes this export process smoother than ever. + To squeeze every drop of performance from these models, export them using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) on NVIDIA hardware or [OpenVINO](https://docs.ultralytics.com/integrations/openvino) for Intel CPUs. The NMS-free design of YOLO26 makes this export process smoother than ever. ## Use Cases and Real-World Applications @@ -97,7 +97,7 @@ Choosing between YOLO11 and YOLO26 largely depends on your specific infrastructu ### Edge Computing and IoT -For applications constrained by power and hardware, such as smart agriculture monitoring via drones or local [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/), **YOLO26** is the undisputed champion. The removal of DFL and the 43% boost in CPU speed means you can run complex vision models on devices without dedicated [GPUs](https://www.nvidia.com/en-us/geforce/graphics-cards/) while maintaining high frame rates. +For applications constrained by power and hardware, such as smart agriculture monitoring via drones or local [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system), **YOLO26** is the undisputed champion. The removal of DFL and the 43% boost in CPU speed means you can run complex vision models on devices without dedicated [GPUs](https://www.nvidia.com/en-us/geforce/graphics-cards/) while maintaining high frame rates. ### Cloud and Enterprise Scale @@ -105,11 +105,11 @@ For applications constrained by power and hardware, such as smart agriculture mo ### Complex Multi-Tasking -If your project requires pinpoint accuracy on tiny objects—such as detecting defects on a circuit board or tracking distant vehicles in [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/)—the **ProgLoss + STAL** implementation in **YOLO26** provides a noticeable uplift in recall and precision for those difficult edge cases. +If your project requires pinpoint accuracy on tiny objects—such as detecting defects on a circuit board or tracking distant vehicles in [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone)—the **ProgLoss + STAL** implementation in **YOLO26** provides a noticeable uplift in recall and precision for those difficult edge cases. ## Training Efficiency and Memory Requirements -A major advantage of the Ultralytics framework is its incredibly low memory footprint during training. Unlike massive vision transformers like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) or the older [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) which can consume vast amounts of CUDA memory, both YOLO11 and YOLO26 are optimized to train efficiently on consumer-grade hardware. +A major advantage of the Ultralytics framework is its incredibly low memory footprint during training. Unlike massive vision transformers like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) or the older [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) which can consume vast amounts of CUDA memory, both YOLO11 and YOLO26 are optimized to train efficiently on consumer-grade hardware. The integration of the MuSGD optimizer in YOLO26 further enhances this by ensuring that the model finds the optimal weights faster, reducing overall GPU compute hours and [cloud computing costs](https://cloud.google.com/products/compute). @@ -134,7 +134,7 @@ model.export(format="onnx") ## Exploring Alternative Architectures -While YOLO26 represents the pinnacle of real-time detection, exploring other models within the Ultralytics documentation can be beneficial. For users tied to legacy environments, earlier architectures like [YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5) still provide robust performance. For zero-shot capabilities where defining classes beforehand isn't possible, [YOLO-World](https://docs.ultralytics.com/models/yolo-world/) offers open-vocabulary detection powered by text prompts. +While YOLO26 represents the pinnacle of real-time detection, exploring other models within the Ultralytics documentation can be beneficial. For users tied to legacy environments, earlier architectures like [YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5) still provide robust performance. For zero-shot capabilities where defining classes beforehand isn't possible, [YOLO-World](https://docs.ultralytics.com/models/yolo-world) offers open-vocabulary detection powered by text prompts. ## Conclusion diff --git a/docs/en/compare/yolo11-vs-yolov10.md b/docs/en/compare/yolo11-vs-yolov10.md index 0c18889b2d3..aca4589497d 100644 --- a/docs/en/compare/yolo11-vs-yolov10.md +++ b/docs/en/compare/yolo11-vs-yolov10.md @@ -6,7 +6,7 @@ keywords: YOLO11, YOLOv10, Ultralytics comparison, object detection models, real # YOLO11 vs YOLOv10: A Comprehensive Technical Comparison of Real-Time Object Detectors -The landscape of real-time computer vision is constantly evolving, with new architectures pushing the boundaries of what is possible on both edge devices and cloud infrastructure. In this detailed technical analysis, we explore the nuances between two pivotal models in the domain: [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv10](https://docs.ultralytics.com/models/yolov10/). Both represent significant leaps in [object detection](https://docs.ultralytics.com/tasks/detect/) capabilities, yet they adopt fundamentally different architectural philosophies to achieve their performance. +The landscape of real-time computer vision is constantly evolving, with new architectures pushing the boundaries of what is possible on both edge devices and cloud infrastructure. In this detailed technical analysis, we explore the nuances between two pivotal models in the domain: [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv10](https://docs.ultralytics.com/models/yolov10). Both represent significant leaps in [object detection](https://docs.ultralytics.com/tasks/detect) capabilities, yet they adopt fundamentally different architectural philosophies to achieve their performance. @@ -21,7 +21,7 @@ The landscape of real-time computer vision is constantly evolving, with new arch - Organization: [Ultralytics](https://www.ultralytics.com/) - Date: 2024-09-27 - GitHub: [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- Docs: [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11/) +- Docs: [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11) Introduced as a versatile powerhouse, YOLO11 builds upon years of foundational research in [computer vision and AI](https://www.ultralytics.com/blog/a-quick-overview-of-vision-ai-and-how-it-works). The core design philosophy of YOLO11 revolves around feature richness and extreme versatility across multiple [computer vision tasks](https://www.ultralytics.com/blog/all-you-need-to-know-about-computer-vision-tasks). @@ -40,17 +40,17 @@ YOLO11 utilizes an anchor-free design that minimizes the complexity of hyperpara - Date: 2024-05-23 - Arxiv: [https://arxiv.org/abs/2405.14458](https://arxiv.org/abs/2405.14458) - GitHub: [https://github.com/THU-MIG/yolov10](https://github.com/THU-MIG/yolov10) -- Docs: [https://docs.ultralytics.com/models/yolov10/](https://docs.ultralytics.com/models/yolov10/) +- Docs: [https://docs.ultralytics.com/models/yolov10/](https://docs.ultralytics.com/models/yolov10) Developed by researchers at Tsinghua University, YOLOv10 made waves as an end-to-end pioneer in the YOLO family. The hallmark of YOLOv10 is its **NMS-Free Training** methodology. By employing consistent dual assignments during the training phase, the model naturally predicts exactly one bounding box per object. This breakthrough completely eliminates the need for [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) during inference, a post-processing step that historically introduced latency bottlenecks in deployment pipelines. The architecture also introduces a holistic efficiency-accuracy design strategy. It incorporates spatial-channel decoupled downsampling and rank-guided block designs that selectively reduce redundancy in the network stages. This results in fewer [FLOPs](https://www.ultralytics.com/glossary/flops) and reduced computational overhead without significantly sacrificing the [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map). For real-time applications where every millisecond counts, the removal of NMS provides a deterministic inference graph highly suitable for [edge AI devices](https://www.ultralytics.com/glossary/edge-ai). -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## Performance Metrics and Benchmarks -When evaluating these two models, we look at a balance of accuracy, parameter count, and speed. The following table showcases how they compare across various scales on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). +When evaluating these two models, we look at a balance of accuracy, parameter count, and speed. The following table showcases how they compare across various scales on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | -------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -67,19 +67,19 @@ When evaluating these two models, we look at a balance of accuracy, parameter co | YOLOv10l | 640 | 53.3 | - | 8.33 | 29.5 | 120.3 | | YOLOv10x | 640 | 54.4 | - | 12.2 | 56.9 | 160.4 | -As observed in the [YOLO performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/), YOLO11 generally achieves slightly higher mAP scores across its variants, particularly in the larger models. The NMS-free design of YOLOv10 ensures highly stable end-to-end inference times, but YOLO11 still manages exceptional throughput when optimized with [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) on NVIDIA hardware. +As observed in the [YOLO performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics), YOLO11 generally achieves slightly higher mAP scores across its variants, particularly in the larger models. The NMS-free design of YOLOv10 ensures highly stable end-to-end inference times, but YOLO11 still manages exceptional throughput when optimized with [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) on NVIDIA hardware. !!! tip "Exporting for Production" - When preparing your models for deployment, exporting to optimized formats is crucial. Both YOLO11 and YOLOv10 can be seamlessly exported to formats like ONNX and TensorRT using the Ultralytics framework. See our guide on [model deployment options](https://docs.ultralytics.com/guides/model-deployment-options/) for step-by-step instructions. + When preparing your models for deployment, exporting to optimized formats is crucial. Both YOLO11 and YOLOv10 can be seamlessly exported to formats like ONNX and TensorRT using the Ultralytics framework. See our guide on [model deployment options](https://docs.ultralytics.com/guides/model-deployment-options) for step-by-step instructions. ## The Ultralytics Ecosystem Advantage While standalone performance metrics are important, the surrounding framework dictates the practical success of a machine learning project. This is where YOLO11, as a native citizen of the Ultralytics ecosystem, truly shines. -The [Ultralytics Platform](https://platform.ultralytics.com) offers an incredibly streamlined user experience. With a simple and unified [Python API](https://docs.ultralytics.com/usage/python/), developers can handle tasks beyond basic bounding boxes. YOLO11 supports native [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection out of the box. This immense versatility is often lacking in specialized research repositories. +The [Ultralytics Platform](https://platform.ultralytics.com) offers an incredibly streamlined user experience. With a simple and unified [Python API](https://docs.ultralytics.com/usage/python), developers can handle tasks beyond basic bounding boxes. YOLO11 supports native [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection out of the box. This immense versatility is often lacking in specialized research repositories. -Furthermore, the ecosystem is backed by extensive documentation and active community support. Integrations with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) for experiment tracking, and [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) for Intel hardware optimization, are built directly into the library. Training a model requires minimal boilerplate code and benefits from highly efficient training processes that require less CUDA memory than heavy transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +Furthermore, the ecosystem is backed by extensive documentation and active community support. Integrations with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) for experiment tracking, and [OpenVINO](https://docs.ultralytics.com/integrations/openvino) for Intel hardware optimization, are built directly into the library. Training a model requires minimal boilerplate code and benefits from highly efficient training processes that require less CUDA memory than heavy transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). ### Hands-On Code Example @@ -110,9 +110,9 @@ Choosing between YOLO11 and YOLOv10 depends on your specific project requirement YOLO11 is a strong choice for: -- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) where reliability and active maintenance are paramount. -- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [OBB](https://docs.ultralytics.com/tasks/obb/) within a single unified framework. -- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python/). +- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) where reliability and active maintenance are paramount. +- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [OBB](https://docs.ultralytics.com/tasks/obb) within a single unified framework. +- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python). ### When to Choose YOLOv10 @@ -124,11 +124,11 @@ YOLOv10 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Next Generation: YOLO26 @@ -144,4 +144,4 @@ YOLO26 also brings Large Language Model (LLM) training stability to computer vis Choosing the right vision model depends on your specific operational constraints. YOLOv10 stands as a significant milestone in academia, proving that NMS can be effectively eliminated from the detection pipeline. However, for a superior balance of performance, comprehensive task versatility, and seamless deployment tools, **YOLO11** offers a robust, enterprise-ready solution. -For engineers who want the absolute cutting edge—combining end-to-end simplicity with blazing-fast edge performance—migrating to the latest [YOLO26](https://docs.ultralytics.com/models/yolo26/) is the ultimate recommendation. By leveraging the comprehensive [Ultralytics Platform](https://platform.ultralytics.com), you ensure your projects are built on a well-maintained, highly efficient, and future-proof foundation. +For engineers who want the absolute cutting edge—combining end-to-end simplicity with blazing-fast edge performance—migrating to the latest [YOLO26](https://docs.ultralytics.com/models/yolo26) is the ultimate recommendation. By leveraging the comprehensive [Ultralytics Platform](https://platform.ultralytics.com), you ensure your projects are built on a well-maintained, highly efficient, and future-proof foundation. diff --git a/docs/en/compare/yolo11-vs-yolov5.md b/docs/en/compare/yolo11-vs-yolov5.md index c95b054afb7..2c8bed1425f 100644 --- a/docs/en/compare/yolo11-vs-yolov5.md +++ b/docs/en/compare/yolo11-vs-yolov5.md @@ -8,7 +8,7 @@ keywords: YOLO11 vs YOLOv5,Yolo comparison,Yolo models,object detection,Yolo per Selecting the right neural network architecture is a pivotal decision for any [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) initiative. As the landscape of [artificial intelligence](https://www.ultralytics.com/glossary/artificial-intelligence-ai) evolves, so do the tools available to developers and researchers. This comprehensive guide provides an in-depth technical comparison between two landmark models from the [Ultralytics](https://www.ultralytics.com/) ecosystem: the highly celebrated YOLOv5 and the advanced YOLO11. -Whether you are deploying lightweight models for [edge AI](https://www.ultralytics.com/glossary/edge-ai) applications or processing high-resolution video streams on cloud GPUs, understanding the architectural nuances, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/), and ideal use cases for these models will ensure you make a data-driven choice for your specific deployment constraints. +Whether you are deploying lightweight models for [edge AI](https://www.ultralytics.com/glossary/edge-ai) applications or processing high-resolution video streams on cloud GPUs, understanding the architectural nuances, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics), and ideal use cases for these models will ensure you make a data-driven choice for your specific deployment constraints. @@ -43,9 +43,9 @@ Both models reflect Ultralytics' commitment to open-source collaboration, robust The evolution from YOLOv5 to YOLO11 introduces several profound architectural shifts designed to optimize accuracy and parameter efficiency. -YOLOv5 was a trailblazer in the [PyTorch](https://pytorch.org/) ecosystem, introducing a highly optimized CSPNet (Cross Stage Partial Network) backbone and a PANet (Path Aggregation Network) neck. It relied on anchor-based detection, which required predefined [anchor boxes](https://www.ultralytics.com/glossary/anchor-boxes) to predict object boundaries. While highly effective, tuning these anchors for custom [computer vision datasets](https://docs.ultralytics.com/datasets/) could be cumbersome. +YOLOv5 was a trailblazer in the [PyTorch](https://pytorch.org/) ecosystem, introducing a highly optimized CSPNet (Cross Stage Partial Network) backbone and a PANet (Path Aggregation Network) neck. It relied on anchor-based detection, which required predefined [anchor boxes](https://www.ultralytics.com/glossary/anchor-boxes) to predict object boundaries. While highly effective, tuning these anchors for custom [computer vision datasets](https://docs.ultralytics.com/datasets) could be cumbersome. -In contrast, YOLO11 transitions to a more modern, anchor-free detection paradigm. This eliminates the need for manual anchor box tuning, streamlining the training process and improving generalization across diverse datasets like the [COCO dataset](https://cocodataset.org/). Additionally, YOLO11 features a decoupled head, meaning classification and bounding box regression tasks are processed in separate branches. This separation significantly improves convergence speed and [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map), particularly for complex [object detection](https://docs.ultralytics.com/tasks/detect/) scenarios. +In contrast, YOLO11 transitions to a more modern, anchor-free detection paradigm. This eliminates the need for manual anchor box tuning, streamlining the training process and improving generalization across diverse datasets like the [COCO dataset](https://cocodataset.org/). Additionally, YOLO11 features a decoupled head, meaning classification and bounding box regression tasks are processed in separate branches. This separation significantly improves convergence speed and [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map), particularly for complex [object detection](https://docs.ultralytics.com/tasks/detect) scenarios. ## Performance Metrics and Benchmarks @@ -73,7 +73,7 @@ A core tenet of the Ultralytics philosophy is exceptional ease of use, supported YOLOv5 historically relied on robust command-line interface (CLI) scripts (`train.py`, `detect.py`) for execution. While powerful, integrating these scripts directly into custom Python applications often required workarounds. -YOLO11 revolutionized this by introducing the streamlined `ultralytics` Python package. This unified API handles everything from training to [exporting models](https://docs.ultralytics.com/modes/export/) formats like [ONNX](https://onnx.ai/), [OpenVINO](https://docs.ultralytics.com/integrations/openvino/), and [TensorRT](https://developer.nvidia.com/tensorrt) natively. +YOLO11 revolutionized this by introducing the streamlined `ultralytics` Python package. This unified API handles everything from training to [exporting models](https://docs.ultralytics.com/modes/export) formats like [ONNX](https://onnx.ai/), [OpenVINO](https://docs.ultralytics.com/integrations/openvino), and [TensorRT](https://developer.nvidia.com/tensorrt) natively. !!! tip "Streamlined Deployment with Ultralytics Platform" @@ -122,7 +122,7 @@ Despite the newer generation, YOLOv5 remains a powerhouse. It is highly recommen YOLO11 represents the ideal choice for modern production pipelines due to its incredible versatility: -- **Multi-Task Environments:** Unlike YOLOv5, which is primarily a detector (with later segmentation additions), YOLO11 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection out of the box. +- **Multi-Task Environments:** Unlike YOLOv5, which is primarily a detector (with later segmentation additions), YOLO11 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection out of the box. - **High-Density Video Analytics:** Ideal for intelligent traffic systems or [retail inventory management](https://www.ultralytics.com/solutions/ai-in-retail) where extracting maximum precision from complex scenes is critical. ## Looking Forward: The YOLO26 Architecture diff --git a/docs/en/compare/yolo11-vs-yolov6.md b/docs/en/compare/yolo11-vs-yolov6.md index 5a3d451da90..b0d88f0dc46 100644 --- a/docs/en/compare/yolo11-vs-yolov6.md +++ b/docs/en/compare/yolo11-vs-yolov6.md @@ -6,7 +6,7 @@ keywords: YOLO11, YOLOv6-3.0, object detection, model comparison, computer visio # YOLO11 vs YOLOv6-3.0: A Comprehensive Technical Comparison -The field of [computer vision](https://en.wikipedia.org/wiki/Computer_vision) evolves rapidly, and selecting the right model architecture is a critical decision for machine learning practitioners. Two significant milestones in the progression of real-time [object detection](https://docs.ultralytics.com/tasks/detect/) are **YOLO11** and **YOLOv6-3.0**. While both models offer impressive capabilities for extracting insights from visual data, they were developed with different primary objectives and design philosophies. +The field of [computer vision](https://en.wikipedia.org/wiki/Computer_vision) evolves rapidly, and selecting the right model architecture is a critical decision for machine learning practitioners. Two significant milestones in the progression of real-time [object detection](https://docs.ultralytics.com/tasks/detect) are **YOLO11** and **YOLOv6-3.0**. While both models offer impressive capabilities for extracting insights from visual data, they were developed with different primary objectives and design philosophies. This guide provides an in-depth technical analysis comparing their architectures, performance metrics, and ideal deployment scenarios to help you make an informed decision for your next AI project. @@ -27,7 +27,7 @@ Developed natively within the Ultralytics ecosystem, YOLO11 was engineered to pr - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2024-09-27 - **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11/) +- **Docs:** [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11) [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } @@ -40,15 +40,15 @@ YOLOv6-3.0 was explicitly tailored for industrial applications where dedicated [ - **Date:** 2023-01-13 - **Arxiv:** [2301.05586](https://arxiv.org/abs/2301.05586) - **GitHub:** [Meituan YOLOv6 Repository](https://github.com/meituan/YOLOv6) -- **Docs:** [YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6/) +- **Docs:** [YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6) -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Architectural Differences The underlying architecture dictates how a model learns and scales. Both frameworks introduce unique enhancements to the classic YOLO formula. -YOLO11 builds upon years of research to deliver an architecture that is incredibly parameter-efficient. It features an advanced backbone and a generalized head capable of handling diverse computer vision tasks—such as [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [pose estimation](https://docs.ultralytics.com/tasks/pose/)—without requiring massive structural overhauls. Furthermore, YOLO11 boasts exceptionally low [CUDA](https://developer.nvidia.com/cuda) memory requirements during training, setting it apart from bulkier [transformer models]() like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +YOLO11 builds upon years of research to deliver an architecture that is incredibly parameter-efficient. It features an advanced backbone and a generalized head capable of handling diverse computer vision tasks—such as [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [pose estimation](https://docs.ultralytics.com/tasks/pose)—without requiring massive structural overhauls. Furthermore, YOLO11 boasts exceptionally low [CUDA](https://developer.nvidia.com/cuda) memory requirements during training, setting it apart from bulkier [transformer models]() like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). Conversely, YOLOv6-3.0 employs a Bi-directional Concatenation (BiC) module and an Anchor-Aided Training (AAT) strategy. These mechanisms are designed to improve localization accuracy. The architecture is primarily decoupled and heavily quantized to favor INT8 [model inference](https://en.wikipedia.org/wiki/Inference), making it a strong contender for high-speed manufacturing lines running legacy GPU stacks. @@ -73,7 +73,7 @@ When evaluating models, [mean Average Precision (mAP)](https://www.ultralytics.c | YOLOv6-3.0m | 640 | 50.0 | - | 5.28 | 34.9 | 85.8 | | YOLOv6-3.0l | 640 | 52.8 | - | 8.95 | 59.6 | 150.7 | -As demonstrated, YOLO11 consistently achieves higher accuracy (mAP) with significantly fewer parameters and FLOPs across equivalent tiers. This parameter efficiency translates directly to lower memory requirements during both [model training](https://docs.ultralytics.com/modes/train/) and inference. +As demonstrated, YOLO11 consistently achieves higher accuracy (mAP) with significantly fewer parameters and FLOPs across equivalent tiers. This parameter efficiency translates directly to lower memory requirements during both [model training](https://docs.ultralytics.com/modes/train) and inference. ## The Ultralytics Advantage @@ -81,7 +81,7 @@ Choosing a model is about more than just raw metrics; it is about the entire [ma 1. **Ease of Use:** The Ultralytics Python API allows you to train, validate, and export models with just a few lines of code. There is no need to manually configure complex dependency trees. 2. **Well-Maintained Ecosystem:** Ultralytics provides a unified ecosystem that receives frequent updates. By utilizing the [Ultralytics Platform](https://platform.ultralytics.com/), developers gain access to collaborative dataset annotation, cloud training, and seamless model monitoring. -3. **Versatility:** Unlike YOLOv6-3.0, which is primarily a bounding box detector, YOLO11 natively supports [image classification](https://docs.ultralytics.com/tasks/classify/) and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), allowing you to consolidate your technology stack. +3. **Versatility:** Unlike YOLOv6-3.0, which is primarily a bounding box detector, YOLO11 natively supports [image classification](https://docs.ultralytics.com/tasks/classify) and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb), allowing you to consolidate your technology stack. 4. **Training Efficiency:** Leveraging modern optimizations and auto-batching, YOLO11 trains efficiently on consumer-grade hardware, democratizing access to state-of-the-art vision AI. ### Code Example: Training and Inference @@ -120,13 +120,13 @@ While YOLO11 represents a massive leap forward, Ultralytics continually pushes t YOLO26 introduces several groundbreaking features designed specifically for modern deployment challenges: -- **End-to-End NMS-Free Design:** Building on concepts pioneered by [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 is natively end-to-end. It completely eliminates Non-Maximum Suppression (NMS) post-processing, resulting in faster, drastically simpler deployment pipelines. +- **End-to-End NMS-Free Design:** Building on concepts pioneered by [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 is natively end-to-end. It completely eliminates Non-Maximum Suppression (NMS) post-processing, resulting in faster, drastically simpler deployment pipelines. - **DFL Removal:** By removing Distribution Focal Loss, YOLO26 simplifies the network head, greatly enhancing compatibility with low-power [Internet of Things (IoT)](https://en.wikipedia.org/wiki/Internet_of_things) and edge devices. - **MuSGD Optimizer:** Inspired by large language model (LLM) training innovations (such as Moonshot AI's Kimi K2), YOLO26 utilizes a hybrid Muon-SGD optimizer, ensuring unmatched training stability and faster convergence. - **Up to 43% Faster CPU Inference:** For applications running without dedicated GPU accelerators, YOLO26 has been heavily optimized for raw CPU throughput. -- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, which is critical for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and aerial surveillance. +- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, which is critical for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and aerial surveillance. - **Task-Specific Improvements:** YOLO26 includes customized enhancements across all tasks, such as multi-scale prototyping for segmentation and Residual Log-Likelihood Estimation (RLE) for pose estimation. If you are starting a new computer vision initiative today, leveraging the [Ultralytics Platform](https://platform.ultralytics.com/) to train a YOLO26 model will ensure your application is built on the most efficient, accurate, and future-proof architecture available. -For developers interested in exploring open-vocabulary detection, you can also review our documentation on [YOLO-World](https://docs.ultralytics.com/models/yolo-world/). +For developers interested in exploring open-vocabulary detection, you can also review our documentation on [YOLO-World](https://docs.ultralytics.com/models/yolo-world). diff --git a/docs/en/compare/yolo11-vs-yolov7.md b/docs/en/compare/yolo11-vs-yolov7.md index 02ded19bce8..cc5180482a2 100644 --- a/docs/en/compare/yolo11-vs-yolov7.md +++ b/docs/en/compare/yolo11-vs-yolov7.md @@ -22,7 +22,7 @@ Authors: Glenn Jocher and Jing Qiu Organization: [Ultralytics](https://www.ultralytics.com) Date: 2024-09-27 GitHub: [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -Docs: [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11/) +Docs: [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11) [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } @@ -32,9 +32,9 @@ Organization: [Institute of Information Science, Academia Sinica, Taiwan](https: Date: 2022-07-06 Arxiv: [https://arxiv.org/abs/2207.02696](https://arxiv.org/abs/2207.02696) GitHub: [https://github.com/WongKinYiu/yolov7](https://github.com/WongKinYiu/yolov7) -Docs: [https://docs.ultralytics.com/models/yolov7/](https://docs.ultralytics.com/models/yolov7/) +Docs: [https://docs.ultralytics.com/models/yolov7/](https://docs.ultralytics.com/models/yolov7) -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ## Architectural Differences @@ -96,11 +96,11 @@ In contrast, a typical YOLOv7 training command looks like this, requiring carefu python train.py --workers 8 --device 0 --batch-size 32 --data data/coco.yaml --img 640 640 --cfg cfg/training/yolov7.yaml --weights 'yolov7_training.pt' ``` -YOLO11 also provides immense versatility. While YOLOv7 requires entirely different codebases or heavy modifications to support tasks beyond detection (like pose or segmentation), YOLO11 handles [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection via a single, cohesive framework. +YOLO11 also provides immense versatility. While YOLOv7 requires entirely different codebases or heavy modifications to support tasks beyond detection (like pose or segmentation), YOLO11 handles [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection via a single, cohesive framework. !!! note "Exporting Made Easy" - Exporting YOLO11 to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) requires just a single command, mitigating the typical operator support issues encountered with legacy models. + Exporting YOLO11 to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino) requires just a single command, mitigating the typical operator support issues encountered with legacy models. ## Real-World Applications and Ideal Use Cases @@ -125,4 +125,4 @@ Released in January 2026, YOLO26 introduces an end-to-end NMS-Free Design, compl [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } -For users interested in specialized alternative structures, exploring the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) or the dynamic open-vocabulary [YOLO-World](https://docs.ultralytics.com/models/yolo-world/) models may also yield beneficial results for diverse computer vision deployments. +For users interested in specialized alternative structures, exploring the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr) or the dynamic open-vocabulary [YOLO-World](https://docs.ultralytics.com/models/yolo-world) models may also yield beneficial results for diverse computer vision deployments. diff --git a/docs/en/compare/yolo11-vs-yolov8.md b/docs/en/compare/yolo11-vs-yolov8.md index 8f563cc584e..fa4f8026534 100644 --- a/docs/en/compare/yolo11-vs-yolov8.md +++ b/docs/en/compare/yolo11-vs-yolov8.md @@ -27,7 +27,7 @@ YOLO11 represents a significant leap forward in optimizing parameter usage. It r - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2024-09-27 - **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- **Documentation:** [YOLO11 Docs](https://docs.ultralytics.com/models/yolo11/) +- **Documentation:** [YOLO11 Docs](https://docs.ultralytics.com/models/yolo11) [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } @@ -39,7 +39,7 @@ Launched a year earlier, YOLOv8 pioneered the transition to an anchor-free detec - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2023-01-10 - **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- **Documentation:** [YOLOv8 Docs](https://docs.ultralytics.com/models/yolov8/) +- **Documentation:** [YOLOv8 Docs](https://docs.ultralytics.com/models/yolov8) [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -79,11 +79,11 @@ The `ultralytics` Python package provides a streamlined API that allows engineer ### Training Efficiency and Memory Requirements -Unlike heavy [Vision Transformers](https://arxiv.org/abs/2010.11929) (like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/)), Ultralytics YOLO models are renowned for their low memory usage during training. This memory efficiency enables developers to train state-of-the-art networks on consumer-grade GPUs or cloud environments like [Google Colab](https://colab.research.google.com/) without facing out-of-memory errors. +Unlike heavy [Vision Transformers](https://arxiv.org/abs/2010.11929) (like [RT-DETR](https://docs.ultralytics.com/models/rtdetr)), Ultralytics YOLO models are renowned for their low memory usage during training. This memory efficiency enables developers to train state-of-the-art networks on consumer-grade GPUs or cloud environments like [Google Colab](https://colab.research.google.com/) without facing out-of-memory errors. ### Versatility Across Vision Tasks -Both YOLO11 and YOLOv8 are true multi-task learners. Beyond standard bounding box [object detection](https://docs.ultralytics.com/tasks/detect/), they natively support [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), human [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) for aerial imagery. +Both YOLO11 and YOLOv8 are true multi-task learners. Beyond standard bounding box [object detection](https://docs.ultralytics.com/tasks/detect), they natively support [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), human [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb) for aerial imagery. ## Use Cases and Recommendations @@ -93,25 +93,25 @@ Choosing between YOLO11 and YOLOv8 depends on your specific project requirements YOLO11 is a strong choice for: -- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) where reliability and active maintenance are paramount. -- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [OBB](https://docs.ultralytics.com/tasks/obb/) within a single unified framework. -- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python/). +- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) where reliability and active maintenance are paramount. +- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [OBB](https://docs.ultralytics.com/tasks/obb) within a single unified framework. +- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python). ### When to Choose YOLOv8 YOLOv8 is recommended for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Code Example: Getting Started @@ -146,9 +146,9 @@ While YOLO11 represents a mature and highly capable architecture, the rapid pace YOLO26 pushes the boundaries of computer vision with several groundbreaking features: -- **End-to-End NMS-Free Design:** Building on concepts explored in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing, resulting in lower, more predictable latency across all deployment hardware. +- **End-to-End NMS-Free Design:** Building on concepts explored in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing, resulting in lower, more predictable latency across all deployment hardware. - **Up to 43% Faster CPU Inference:** By completely removing the Distribution Focal Loss (DFL) branch, YOLO26 is specifically optimized for [edge computing devices](https://www.ultralytics.com/glossary/edge-ai) that lack powerful GPUs. - **MuSGD Optimizer:** Inspired by large language model (LLM) training techniques, YOLO26 utilizes a hybrid MuSGD optimizer, ensuring remarkably stable and rapid training convergence. - **ProgLoss + STAL:** These advanced loss functions yield notable improvements in tiny and heavily occluded object recognition, essential for autonomous robotics and drone-based analytics. -Whether you rely on the proven reliability of YOLOv8, the optimized architecture of YOLO11, or the next-generation capabilities of YOLO26, the [Ultralytics Platform](https://platform.ultralytics.com) ensures you have the tools necessary to bring your vision AI applications from concept to production seamlessly. Ensure you explore the extensive [integrations](https://docs.ultralytics.com/integrations/) available to connect your models with enterprise workflows and analytics dashboards. +Whether you rely on the proven reliability of YOLOv8, the optimized architecture of YOLO11, or the next-generation capabilities of YOLO26, the [Ultralytics Platform](https://platform.ultralytics.com) ensures you have the tools necessary to bring your vision AI applications from concept to production seamlessly. Ensure you explore the extensive [integrations](https://docs.ultralytics.com/integrations) available to connect your models with enterprise workflows and analytics dashboards. diff --git a/docs/en/compare/yolo11-vs-yolov9.md b/docs/en/compare/yolo11-vs-yolov9.md index d9808a35db5..49c0ddd7f21 100644 --- a/docs/en/compare/yolo11-vs-yolov9.md +++ b/docs/en/compare/yolo11-vs-yolov9.md @@ -25,7 +25,7 @@ YOLO11 is a highly optimized, versatile model designed for production-grade envi - **Organization:** [Ultralytics](https://www.ultralytics.com) - **Date:** 2024-09-27 - **GitHub:** [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11/) +- **Docs:** [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11) [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } @@ -38,9 +38,9 @@ YOLOv9 is a powerful academic contribution that introduces novel concepts to mit - **Date:** 2024-02-21 - **Arxiv:** [https://arxiv.org/abs/2402.13616](https://arxiv.org/abs/2402.13616) - **GitHub:** [https://github.com/WongKinYiu/yolov9](https://github.com/WongKinYiu/yolov9) -- **Docs:** [https://docs.ultralytics.com/models/yolov9/](https://docs.ultralytics.com/models/yolov9/) +- **Docs:** [https://docs.ultralytics.com/models/yolov9/](https://docs.ultralytics.com/models/yolov9) -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ## Architectural Innovations @@ -50,7 +50,7 @@ YOLOv9 tackles the "information bottleneck" problem—where data is lost as it p ### YOLO11: Ecosystem and Efficiency -While YOLOv9 focuses on gradient flow, YOLO11 is engineered for real-world robustness and versatility. It refines the fundamental YOLO architecture to drastically reduce CUDA memory requirements during training compared to transformer-heavy alternatives. Furthermore, YOLO11 is not just an object detector; it natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +While YOLOv9 focuses on gradient flow, YOLO11 is engineered for real-world robustness and versatility. It refines the fundamental YOLO architecture to drastically reduce CUDA memory requirements during training compared to transformer-heavy alternatives. Furthermore, YOLO11 is not just an object detector; it natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb). !!! tip "Streamlined Development" @@ -111,7 +111,7 @@ YOLOv9 is a fantastic choice for researchers looking to explore deep learning ar ### When to Choose YOLO11 -YOLO11 is the ultimate tool for production. Its streamlined [object detection](https://docs.ultralytics.com/tasks/detect/) capabilities make it perfect for [smart city traffic management](https://en.wikipedia.org/wiki/Smart_city) and edge devices like the Raspberry Pi or NVIDIA Jetson. Furthermore, its versatility across various tasks means a single development pipeline can handle [segmentation in manufacturing](https://www.ultralytics.com/blog/improving-manufacturing-with-computer-vision) and [pose estimation in sports analytics](https://www.ultralytics.com/blog/using-pose-estimation-to-perfect-your-running-technique). +YOLO11 is the ultimate tool for production. Its streamlined [object detection](https://docs.ultralytics.com/tasks/detect) capabilities make it perfect for [smart city traffic management](https://en.wikipedia.org/wiki/Smart_city) and edge devices like the Raspberry Pi or NVIDIA Jetson. Furthermore, its versatility across various tasks means a single development pipeline can handle [segmentation in manufacturing](https://www.ultralytics.com/blog/improving-manufacturing-with-computer-vision) and [pose estimation in sports analytics](https://www.ultralytics.com/blog/using-pose-estimation-to-perfect-your-running-technique). ## The Cutting Edge: Enter YOLO26 @@ -125,7 +125,7 @@ YOLO26 combines the best of recent innovations into a production-ready powerhous - **Up to 43% Faster CPU Inference:** Specifically optimized for edge computing devices without dedicated GPUs. - **ProgLoss + STAL:** These improved loss functions drastically enhance small-object recognition, which is critical for [agricultural monitoring](https://www.ultralytics.com/blog/the-changing-landscape-of-ai-in-agriculture) and aerial imagery. -Users interested in exploring diverse architectures might also want to look into [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) for transformer-based tracking or [YOLO-World](https://docs.ultralytics.com/models/yolo-world/) for zero-shot open-vocabulary detection. +Users interested in exploring diverse architectures might also want to look into [RT-DETR](https://docs.ultralytics.com/models/rtdetr) for transformer-based tracking or [YOLO-World](https://docs.ultralytics.com/models/yolo-world) for zero-shot open-vocabulary detection. ## Conclusion diff --git a/docs/en/compare/yolo11-vs-yolox.md b/docs/en/compare/yolo11-vs-yolox.md index 58b2b75b14f..c56876e567a 100644 --- a/docs/en/compare/yolo11-vs-yolox.md +++ b/docs/en/compare/yolo11-vs-yolox.md @@ -25,9 +25,9 @@ Released in September 2024 by Glenn Jocher and Jing Qiu at [Ultralytics](https:/ - **Organization:** Ultralytics - **Date:** 2024-09-27 - **GitHub:** [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11/) +- **Docs:** [https://docs.ultralytics.com/models/yolo11/](https://docs.ultralytics.com/models/yolo11) -YOLO11 goes beyond standard bounding boxes, natively supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. Its refined architecture optimizes feature extraction to ensure better feature retention across complex spatial hierarchies. +YOLO11 goes beyond standard bounding boxes, natively supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. Its refined architecture optimizes feature extraction to ensure better feature retention across complex spatial hierarchies. [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } @@ -52,7 +52,7 @@ YOLOX introduced a decoupled head and an anchor-free paradigm, which significant ## Performance and Metrics -When evaluating detection models, examining the balance of parameters, computational cost (FLOPs), and mean Average Precision (mAP) is crucial for real-world [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/). +When evaluating detection models, examining the balance of parameters, computational cost (FLOPs), and mean Average Precision (mAP) is crucial for real-world [model deployment](https://docs.ultralytics.com/guides/model-deployment-options). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | --------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -75,9 +75,9 @@ As seen in the table, **YOLO11x** significantly outperforms **YOLOXx** in absolu ### The Ultralytics Advantage -One of the most profound differences between YOLO11 and YOLOX lies in usability. YOLOX operates primarily as a research codebase, requiring complex environment configuration, manual compilation of C++ operators, and verbose command-line arguments to initiate [custom dataset training](https://docs.ultralytics.com/guides/custom-trainer/). +One of the most profound differences between YOLO11 and YOLOX lies in usability. YOLOX operates primarily as a research codebase, requiring complex environment configuration, manual compilation of C++ operators, and verbose command-line arguments to initiate [custom dataset training](https://docs.ultralytics.com/guides/custom-trainer). -In stark contrast, YOLO11 is fully integrated into the Ultralytics Python package, providing a streamlined, "zero-to-hero" workflow. The [Ultralytics Platform](https://docs.ultralytics.com/platform/) offers extensive tools for data annotation, experiment tracking, and cloud-based training, abstracting away the boilerplate so engineers can focus on model performance. +In stark contrast, YOLO11 is fully integrated into the Ultralytics Python package, providing a streamlined, "zero-to-hero" workflow. The [Ultralytics Platform](https://docs.ultralytics.com/platform) offers extensive tools for data annotation, experiment tracking, and cloud-based training, abstracting away the boilerplate so engineers can focus on model performance. ```python from ultralytics import YOLO @@ -92,7 +92,7 @@ results = model.train(data="coco8.yaml", epochs=100, imgsz=640) model.export(format="onnx") ``` -Furthermore, exporting an Ultralytics model to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), CoreML, or [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) requires only a single command, whereas legacy repositories often mandate complex third-party tools or manual graph surgeries. +Furthermore, exporting an Ultralytics model to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), CoreML, or [OpenVINO](https://docs.ultralytics.com/integrations/openvino) requires only a single command, whereas legacy repositories often mandate complex third-party tools or manual graph surgeries. ## Real-World Use Cases @@ -106,7 +106,7 @@ For nearly all modern production scenarios, YOLO11 provides a far superior exper - **Smart Cities and Retail:** Due to its exceptional speed-to-accuracy ratio, YOLO11 handles crowded scenes effortlessly, powering [automated retail analytics](https://www.ultralytics.com/blog/ai-in-retail-enhancing-customer-experience-using-computer-vision) and traffic management systems without requiring massive GPU clusters. - **Edge Computing:** The high memory efficiency and robust export options make YOLO11 perfect for [edge AI deployments](https://www.ultralytics.com/glossary/edge-ai) on devices like Raspberry Pi or NVIDIA Jetson platforms. -- **Complex Pipelines:** If a project demands combining object detection with [pose keypoints](https://docs.ultralytics.com/tasks/pose/) (e.g., sports analytics) or precise instance segmentation (e.g., medical imaging), YOLO11 handles all tasks natively through one unified API. +- **Complex Pipelines:** If a project demands combining object detection with [pose keypoints](https://docs.ultralytics.com/tasks/pose) (e.g., sports analytics) or precise instance segmentation (e.g., medical imaging), YOLO11 handles all tasks natively through one unified API. ## Use Cases and Recommendations @@ -116,9 +116,9 @@ Choosing between YOLO11 and YOLOX depends on your specific project requirements, YOLO11 is a strong choice for: -- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) where reliability and active maintenance are paramount. -- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [OBB](https://docs.ultralytics.com/tasks/obb/) within a single unified framework. -- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python/). +- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) where reliability and active maintenance are paramount. +- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [OBB](https://docs.ultralytics.com/tasks/obb) within a single unified framework. +- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python). ### When to Choose YOLOX @@ -130,22 +130,22 @@ YOLOX is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Looking Ahead: The Power of YOLO26 While YOLO11 stands as an exceptional choice, the landscape of AI continually accelerates. For teams seeking the absolute pinnacle of efficiency and stability, **[YOLO26](https://platform.ultralytics.com/ultralytics/yolo26)** (released January 2026) is the ultimate recommendation for new computer vision projects. -YOLO26 represents a massive leap forward by implementing an **End-to-End NMS-Free Design**. By eliminating [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing, it completely removes latency variability, dramatically simplifying deployment logic—a concept first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/). +YOLO26 represents a massive leap forward by implementing an **End-to-End NMS-Free Design**. By eliminating [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing, it completely removes latency variability, dramatically simplifying deployment logic—a concept first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10). -Furthermore, YOLO26 features **DFL Removal** (Distribution Focal Loss), optimizing the architecture to achieve up to **43% faster CPU inference**, making it the undisputed champion for low-power and edge devices. Training stability is also supercharged via the **MuSGD Optimizer**—an LLM-inspired hybrid of SGD and Muon that accelerates convergence. Combined with advanced loss functions like **ProgLoss + STAL**, YOLO26 excels at detecting small objects in challenging environments like [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and IoT edge sensors. +Furthermore, YOLO26 features **DFL Removal** (Distribution Focal Loss), optimizing the architecture to achieve up to **43% faster CPU inference**, making it the undisputed champion for low-power and edge devices. Training stability is also supercharged via the **MuSGD Optimizer**—an LLM-inspired hybrid of SGD and Muon that accelerates convergence. Combined with advanced loss functions like **ProgLoss + STAL**, YOLO26 excels at detecting small objects in challenging environments like [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and IoT edge sensors. !!! tip "Further Exploration" - Looking to expand your knowledge of object detection architectures? Explore the open-vocabulary capabilities of [YOLO-World](https://docs.ultralytics.com/models/yolo-world/) or dive into the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) model documented in the Ultralytics ecosystem. + Looking to expand your knowledge of object detection architectures? Explore the open-vocabulary capabilities of [YOLO-World](https://docs.ultralytics.com/models/yolo-world) or dive into the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr) model documented in the Ultralytics ecosystem. In conclusion, while YOLOX introduced important architectural concepts in 2021, the comprehensive toolset, memory efficiency, and cutting-edge performance of YOLO11—and especially the revolutionary architecture of YOLO26—make the Ultralytics ecosystem the clear choice for researchers and enterprise developers today. diff --git a/docs/en/compare/yolo26-vs-damo-yolo.md b/docs/en/compare/yolo26-vs-damo-yolo.md index 02f8e807941..c85e88e393e 100644 --- a/docs/en/compare/yolo26-vs-damo-yolo.md +++ b/docs/en/compare/yolo26-vs-damo-yolo.md @@ -21,7 +21,7 @@ Developed by Glenn Jocher and Jing Qiu at [Ultralytics](https://www.ultralytics. Key architectural breakthroughs of YOLO26 include: -- **End-to-End NMS-Free Design:** Building on pioneering work from [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 is natively end-to-end. By completely eliminating Non-Maximum Suppression (NMS) during post-processing, it guarantees deterministic latency and massively simplifies deployment pipelines. +- **End-to-End NMS-Free Design:** Building on pioneering work from [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 is natively end-to-end. By completely eliminating Non-Maximum Suppression (NMS) during post-processing, it guarantees deterministic latency and massively simplifies deployment pipelines. - **DFL Removal:** The removal of Distribution Focal Loss streamlines the model graph. This makes exporting to deployment frameworks like [ONNX](https://onnx.ai/) and [TensorRT](https://developer.nvidia.com/tensorrt) much smoother and ensures better compatibility with low-power edge devices. - **MuSGD Optimizer:** Inspired by Moonshot AI's Kimi K2, this hybrid of Stochastic Gradient Descent (SGD) and Muon brings LLM training innovations into computer vision, resulting in remarkably stable training and rapid convergence. - **ProgLoss + STAL:** These advanced loss functions deliver notable improvements in small-object recognition, which is a critical necessity for drone-based [aerial imagery analysis](https://www.ultralytics.com/blog/using-computer-vision-to-analyze-satellite-imagery) and intricate robotics pipelines. @@ -73,14 +73,14 @@ The true strength of a machine learning model lies not just in its raw metrics, ### The Ultralytics Advantage -Choosing an Ultralytics model guarantees access to a highly refined, developer-centric ecosystem. Complex workflows involving [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation/), [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/), and robust experiment tracking are abstracted into intuitive commands. +Choosing an Ultralytics model guarantees access to a highly refined, developer-centric ecosystem. Complex workflows involving [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation), [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning), and robust experiment tracking are abstracted into intuitive commands. Furthermore, YOLO26 offers unmatched versatility. While DAMO-YOLO is strictly an object detector, YOLO26 provides comprehensive, task-specific improvements across multiple domains out-of-the-box: -- **[Instance Segmentation](https://docs.ultralytics.com/tasks/segment/):** Utilizing specialized semantic segmentation loss and multi-scale prototyping. -- **[Pose Estimation](https://docs.ultralytics.com/tasks/pose/):** Benefiting from advanced Residual Log-Likelihood Estimation (RLE). -- **[Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/):** Incorporating specialized angle loss functions to perfectly resolve tricky boundary issues. -- **[Image Classification](https://docs.ultralytics.com/tasks/classify/):** For rapid and lightweight global image labeling. +- **[Instance Segmentation](https://docs.ultralytics.com/tasks/segment):** Utilizing specialized semantic segmentation loss and multi-scale prototyping. +- **[Pose Estimation](https://docs.ultralytics.com/tasks/pose):** Benefiting from advanced Residual Log-Likelihood Estimation (RLE). +- **[Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb):** Incorporating specialized angle loss functions to perfectly resolve tricky boundary issues. +- **[Image Classification](https://docs.ultralytics.com/tasks/classify):** For rapid and lightweight global image labeling. ### Training Methodologies @@ -103,7 +103,7 @@ model.export(format="onnx") !!! note "Exploring Other Models" - If you are interested in exploring other modern architectures within the Ultralytics ecosystem, the highly capable [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a fantastic choice for legacy pipelines. Alternatively, researchers interested in transformer-based architectures can explore the [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) model. + If you are interested in exploring other modern architectures within the Ultralytics ecosystem, the highly capable [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a fantastic choice for legacy pipelines. Alternatively, researchers interested in transformer-based architectures can explore the [RT-DETR](https://docs.ultralytics.com/models/rtdetr) model. ## Real-World Applications @@ -111,7 +111,7 @@ Choosing between these architectures ultimately depends on your deployment envir ### Edge AI and IoT Devices -For smart retail cameras, automated agricultural monitors, or [robotics](https://www.ultralytics.com/solutions/ai-in-robotics), compute resources are strictly limited. Here, **YOLO26** is the definitive choice. Its 43% faster CPU inference, completely NMS-free pipeline, and tiny parameter footprint allow it to run smoothly on edge devices like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) without sacrificing critical accuracy. +For smart retail cameras, automated agricultural monitors, or [robotics](https://www.ultralytics.com/solutions/ai-in-robotics), compute resources are strictly limited. Here, **YOLO26** is the definitive choice. Its 43% faster CPU inference, completely NMS-free pipeline, and tiny parameter footprint allow it to run smoothly on edge devices like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) without sacrificing critical accuracy. ### High-Speed Manufacturing and Quality Control @@ -131,7 +131,7 @@ YOLO26 is a strong choice for: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ### When to Choose DAMO-YOLO diff --git a/docs/en/compare/yolo26-vs-efficientdet.md b/docs/en/compare/yolo26-vs-efficientdet.md index 630e87f8d59..d2a57842d8b 100644 --- a/docs/en/compare/yolo26-vs-efficientdet.md +++ b/docs/en/compare/yolo26-vs-efficientdet.md @@ -21,9 +21,9 @@ Whether your deployment targets high-throughput cloud servers or latency-constra **Organization:** [Ultralytics](https://www.ultralytics.com/) **Date:** 2026-01-14 **GitHub:** [Ultralytics GitHub](https://github.com/ultralytics/ultralytics) -**Docs:** [YOLO26 Official Documentation](https://docs.ultralytics.com/models/yolo26/) +**Docs:** [YOLO26 Official Documentation](https://docs.ultralytics.com/models/yolo26) -Released in early 2026, [YOLO26](https://docs.ultralytics.com/models/yolo26/) represents the latest evolution in the YOLO family, specifically engineered to provide an unparalleled user experience and top-tier [mean Average Precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics/). Designed from the ground up for modern hardware, it offers exceptional versatility across [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/). +Released in early 2026, [YOLO26](https://docs.ultralytics.com/models/yolo26) represents the latest evolution in the YOLO family, specifically engineered to provide an unparalleled user experience and top-tier [mean Average Precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics). Designed from the ground up for modern hardware, it offers exceptional versatility across [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose). YOLO26 introduces several groundbreaking features that drastically improve both training stability and inference speeds: @@ -35,7 +35,7 @@ YOLO26 introduces several groundbreaking features that drastically improve both !!! tip "Streamlined Exporting" - Thanks to the DFL removal and NMS-free architecture, exporting YOLO26 models to edge-friendly formats like [NVIDIA TensorRT](https://developer.nvidia.com/tensorrt) or [Intel OpenVINO](https://docs.ultralytics.com/integrations/openvino/) requires virtually no custom plugin development. + Thanks to the DFL removal and NMS-free architecture, exporting YOLO26 models to edge-friendly formats like [NVIDIA TensorRT](https://developer.nvidia.com/tensorrt) or [Intel OpenVINO](https://docs.ultralytics.com/integrations/openvino) requires virtually no custom plugin development. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -54,7 +54,7 @@ Key innovations of EfficientDet include: - **BiFPN (Bi-directional Feature Pyramid Network):** A mechanism that allows easy and fast multi-scale feature fusion, enabling the network to better understand objects of varying sizes. - **Compound Scaling:** A heuristic method to scale up resolution, depth, and width uniformly, creating a family of models from d0 (smallest) to d7 (largest). -While EfficientDet remains a robust choice for strict bounding box detection, it generally lacks the modern multi-task versatility (such as native [OBB tasks](https://docs.ultralytics.com/tasks/obb/)) and the streamlined, unified [Python](https://www.python.org/) ecosystem that modern developers expect. +While EfficientDet remains a robust choice for strict bounding box detection, it generally lacks the modern multi-task versatility (such as native [OBB tasks](https://docs.ultralytics.com/tasks/obb)) and the streamlined, unified [Python](https://www.python.org/) ecosystem that modern developers expect. [Learn more about EfficientDet](https://github.com/google/automl/tree/master/efficientdet#readme){ .md-button } @@ -83,7 +83,7 @@ As shown above, YOLO26 establishes a superior performance balance. The YOLO26x m ## Training Efficiency and The Ecosystem Advantage -A major distinction between the two architectures lies in their development environments. EfficientDet is deeply embedded within the Google AutoML and TensorFlow ecosystem, which, while powerful, can introduce steep learning curves and rigid configurations for custom datasets like [DOTAv1](https://docs.ultralytics.com/datasets/obb/dota-v2/). +A major distinction between the two architectures lies in their development environments. EfficientDet is deeply embedded within the Google AutoML and TensorFlow ecosystem, which, while powerful, can introduce steep learning curves and rigid configurations for custom datasets like [DOTAv1](https://docs.ultralytics.com/datasets/obb/dota-v2). Conversely, Ultralytics offers an incredibly well-maintained ecosystem built on [PyTorch](https://pytorch.org/). The memory usage during training is strictly optimized, allowing engineers to train robust models without requiring excessive VRAM allocations common in transformer-based networks. @@ -118,7 +118,7 @@ model.export(format="engine") **When to use YOLO26:** - **Edge Computing & Mobile:** With up to 43% faster CPU inference and no NMS overhead, YOLO26 excels on devices with strictly constrained compute budgets like Raspberry Pis or mobile phones. -- **Multitasking:** When a single pipeline requires bounding boxes, [segmentation masks](https://docs.ultralytics.com/tasks/segment/), and tracking, the versatility of YOLO26 is unmatched. +- **Multitasking:** When a single pipeline requires bounding boxes, [segmentation masks](https://docs.ultralytics.com/tasks/segment), and tracking, the versatility of YOLO26 is unmatched. - **Drone & Aerial Imagery:** The combination of ProgLoss and STAL greatly enhances the detection of extremely small objects from high altitudes. **When to use EfficientDet:** @@ -128,6 +128,6 @@ model.export(format="engine") ## Exploring Other Alternatives -While this guide focuses heavily on the [YOLO26 vs EfficientDet](https://docs.ultralytics.com/compare/efficientdet-vs-yolo26/) paradigm, the broader Ultralytics ecosystem houses other incredible architectures. If your application relies heavily on transformers, [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) offers real-time transformer-based detection. Alternatively, if you are supporting legacy systems, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains fully supported and highly effective. For a broader overview, visit the [Ultralytics Model Comparisons Hub](https://docs.ultralytics.com/compare/). +While this guide focuses heavily on the [YOLO26 vs EfficientDet](https://docs.ultralytics.com/compare/efficientdet-vs-yolo26) paradigm, the broader Ultralytics ecosystem houses other incredible architectures. If your application relies heavily on transformers, [RT-DETR](https://docs.ultralytics.com/models/rtdetr) offers real-time transformer-based detection. Alternatively, if you are supporting legacy systems, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains fully supported and highly effective. For a broader overview, visit the [Ultralytics Model Comparisons Hub](https://docs.ultralytics.com/compare). Ultimately, for any modern computer vision pipeline built today, the sheer speed, ease of use, and state-of-the-art accuracy of **YOLO26** make it the undisputed recommendation for researchers and developers alike. diff --git a/docs/en/compare/yolo26-vs-pp-yoloe.md b/docs/en/compare/yolo26-vs-pp-yoloe.md index 87007c5ff66..bdcd60c7fc2 100644 --- a/docs/en/compare/yolo26-vs-pp-yoloe.md +++ b/docs/en/compare/yolo26-vs-pp-yoloe.md @@ -24,7 +24,7 @@ Released in January 2026, YOLO26 represents the pinnacle of the Ultralytics ecos - Organization: [Ultralytics](https://www.ultralytics.com) - Date: 2026-01-14 - GitHub: [Ultralytics GitHub Repository](https://github.com/ultralytics/ultralytics) -- Docs: [Official YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26/) +- Docs: [Official YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26) [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -46,12 +46,12 @@ The differences in how these models process visual data drastically impact their ### YOLO26: The NMS-Free Frontier -YOLO26 introduces several breakthrough architectural changes designed for streamlined [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/): +YOLO26 introduces several breakthrough architectural changes designed for streamlined [model deployment](https://docs.ultralytics.com/guides/model-deployment-options): -- **End-to-End NMS-Free Design:** Building on concepts first introduced in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing. This reduces latency variability and massively simplifies deployment pipelines. -- **DFL Removal:** By removing Distribution Focal Loss (DFL), the model is exceptionally lighter, enabling seamless export to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) and [CoreML](https://docs.ultralytics.com/integrations/coreml/). +- **End-to-End NMS-Free Design:** Building on concepts first introduced in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing. This reduces latency variability and massively simplifies deployment pipelines. +- **DFL Removal:** By removing Distribution Focal Loss (DFL), the model is exceptionally lighter, enabling seamless export to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) and [CoreML](https://docs.ultralytics.com/integrations/coreml). - **MuSGD Optimizer:** Inspired by Moonshot AI’s Kimi K2, YOLO26 brings LLM training innovations to computer vision. The hybrid MuSGD optimizer (SGD + Muon) ensures highly stable training dynamics and rapid convergence. -- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, making the architecture highly effective for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and [agricultural applications](https://www.ultralytics.com/blog/computer-vision-in-agriculture-transforming-fruit-detection-and-precision-farming). +- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, making the architecture highly effective for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and [agricultural applications](https://www.ultralytics.com/blog/computer-vision-in-agriculture-transforming-fruit-detection-and-precision-farming). ### PP-YOLOE+: A Paddle-Centric Approach @@ -86,7 +86,7 @@ Thanks to specific edge optimizations and DFL removal, YOLO26 delivers up to **4 While PP-YOLOE+ is a capable model, the true differentiator lies in the developer experience. The integrated [Ultralytics ecosystem](https://platform.ultralytics.com/) provides an unmatched environment for vision AI practitioners. 1. **Ease of Use:** Ultralytics offers a streamlined user experience. A simple Python API abstracts the complexity of data pipelines and training loops, supported by extensive and actively maintained documentation. -2. **Versatility:** Unlike PP-YOLOE+, which is primarily focused on object detection, YOLO26 supports [image classification](https://docs.ultralytics.com/tasks/classify/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) natively using the same API structure. +2. **Versatility:** Unlike PP-YOLOE+, which is primarily focused on object detection, YOLO26 supports [image classification](https://docs.ultralytics.com/tasks/classify), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb) natively using the same API structure. 3. **Training Efficiency:** The automated downloading of readily available pre-trained weights, coupled with advanced augmentations, ensures efficient training processes that require less CUDA memory and time compared to traditional frameworks. ### Code Example: Simplicity in Action @@ -134,7 +134,7 @@ YOLO26 is a strong choice for: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ### When to Choose PP-YOLOE+ @@ -146,6 +146,6 @@ PP-YOLOE+ is recommended for: ## Exploring Other Architectures -For users exploring a broader spectrum of models, we also recommend reviewing [YOLO11](https://docs.ultralytics.com/models/yolo11/), the highly reliable prior generation of Ultralytics models, which remains a staple in thousands of production environments. Additionally, for scenarios requiring transformer-based mechanisms, the [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) architecture offers an intriguing alternative, albeit with higher memory demands during training. +For users exploring a broader spectrum of models, we also recommend reviewing [YOLO11](https://docs.ultralytics.com/models/yolo11), the highly reliable prior generation of Ultralytics models, which remains a staple in thousands of production environments. Additionally, for scenarios requiring transformer-based mechanisms, the [RT-DETR](https://docs.ultralytics.com/models/rtdetr) architecture offers an intriguing alternative, albeit with higher memory demands during training. Ultimately, by leveraging the MuSGD optimizer, ProgLoss + STAL capabilities, and an NMS-free design, YOLO26 cements its position as the premier choice for modern, scalable, and highly efficient vision AI solutions. diff --git a/docs/en/compare/yolo26-vs-rtdetr.md b/docs/en/compare/yolo26-vs-rtdetr.md index 36ebb72cd72..c78cb0642c8 100644 --- a/docs/en/compare/yolo26-vs-rtdetr.md +++ b/docs/en/compare/yolo26-vs-rtdetr.md @@ -23,27 +23,27 @@ Developed by Ultralytics, YOLO26 represents a massive generational leap for the - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2026-01-14 - **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO26 Official Documentation](https://docs.ultralytics.com/models/yolo26/) +- **Docs:** [YOLO26 Official Documentation](https://docs.ultralytics.com/models/yolo26) ### Architectural Innovations and Strengths -YOLO26 introduces several groundbreaking features that differentiate it not only from Transformer models but also from earlier iterations like [YOLO11](https://docs.ultralytics.com/models/yolo11/): +YOLO26 introduces several groundbreaking features that differentiate it not only from Transformer models but also from earlier iterations like [YOLO11](https://docs.ultralytics.com/models/yolo11): -- **End-to-End NMS-Free Design:** YOLO26 eliminates traditional Non-Maximum Suppression (NMS) during post-processing. Pioneered in models like [YOLOv10](https://docs.ultralytics.com/models/yolov10/), this natively end-to-end approach reduces inference latency variance and simplifies deployment logic, particularly on edge hardware. -- **Up to 43% Faster CPU Inference:** Recognizing the growing need for decentralized AI, YOLO26 is highly optimized for devices lacking dedicated GPUs, such as the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/). +- **End-to-End NMS-Free Design:** YOLO26 eliminates traditional Non-Maximum Suppression (NMS) during post-processing. Pioneered in models like [YOLOv10](https://docs.ultralytics.com/models/yolov10), this natively end-to-end approach reduces inference latency variance and simplifies deployment logic, particularly on edge hardware. +- **Up to 43% Faster CPU Inference:** Recognizing the growing need for decentralized AI, YOLO26 is highly optimized for devices lacking dedicated GPUs, such as the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi). - **DFL Removal:** By stripping out the Distribution Focal Loss (DFL), YOLO26 offers a simplified export process and vastly improved compatibility with low-power edge devices and microcontrollers. - **MuSGD Optimizer:** Bridging the gap between Large Language Model (LLM) training and computer vision, YOLO26 utilizes the MuSGD optimizer. This hybrid of SGD and Muon—inspired by Moonshot AI's Kimi K2—ensures robust training stability and faster convergence. -- **ProgLoss + STAL:** Advanced loss functions bring notable improvements to small-object recognition. This is critical for industries relying on [aerial imagery analysis](https://docs.ultralytics.com/datasets/detect/visdrone/) and Internet of Things (IoT) sensors. +- **ProgLoss + STAL:** Advanced loss functions bring notable improvements to small-object recognition. This is critical for industries relying on [aerial imagery analysis](https://docs.ultralytics.com/datasets/detect/visdrone) and Internet of Things (IoT) sensors. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } ### Versatility Across Vision Tasks -Unlike models limited strictly to bounding boxes, YOLO26 is a versatile powerhouse. It incorporates task-specific improvements, such as semantic segmentation loss and multi-scale proto for [instance segmentation](https://docs.ultralytics.com/tasks/segment/), Residual Log-Likelihood Estimation (RLE) for [pose estimation](https://docs.ultralytics.com/tasks/pose/), and specialized angle loss to resolve boundary issues in [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) tasks. +Unlike models limited strictly to bounding boxes, YOLO26 is a versatile powerhouse. It incorporates task-specific improvements, such as semantic segmentation loss and multi-scale proto for [instance segmentation](https://docs.ultralytics.com/tasks/segment), Residual Log-Likelihood Estimation (RLE) for [pose estimation](https://docs.ultralytics.com/tasks/pose), and specialized angle loss to resolve boundary issues in [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) tasks. !!! tip "Edge Deployment Strategy" - When deploying to edge devices, utilize the `YOLO26n` (Nano) or `YOLO26s` (Small) variants. Exporting these models to [CoreML](https://docs.ultralytics.com/integrations/coreml/) or [TFLite](https://docs.ultralytics.com/integrations/tflite/) is frictionless thanks to the DFL removal and NMS-free architecture, guaranteeing smooth real-time performance on iOS and Android. + When deploying to edge devices, utilize the `YOLO26n` (Nano) or `YOLO26s` (Small) variants. Exporting these models to [CoreML](https://docs.ultralytics.com/integrations/coreml) or [TFLite](https://docs.ultralytics.com/integrations/tflite) is frictionless thanks to the DFL removal and NMS-free architecture, guaranteeing smooth real-time performance on iOS and Android. ## RTDETRv2: Enhancing Real-Time Detection Transformers @@ -63,7 +63,7 @@ RTDETRv2 employs a Transformer-based architecture, which inherently processes im - **Bag-of-Freebies:** The v2 iteration introduces a series of optimized training techniques (bag-of-freebies) that improve the baseline performance without adding inference cost. - **Global Context Awareness:** Because of the Transformer attention layers, RTDETRv2 is naturally adept at understanding complex scenes where global context is necessary to distinguish overlapping or occluded objects. -[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr){ .md-button } ### Limitations of Transformer Models @@ -98,7 +98,7 @@ YOLO26 is a strong choice for: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ### When to Choose RT-DETR @@ -114,7 +114,7 @@ Choosing the right machine learning architecture is only part of the equation; t ### Ease of Use and Training Efficiency -The [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) offers a remarkably streamlined experience. Training complex models no longer requires verbose boilerplate code. Furthermore, YOLO26's training efficiency is substantially better, utilizing far less GPU VRAM than the memory-intensive attention mechanisms of RTDETRv2, allowing for larger batch sizes even on consumer-grade hardware. +The [Ultralytics Python API](https://docs.ultralytics.com/usage/python) offers a remarkably streamlined experience. Training complex models no longer requires verbose boilerplate code. Furthermore, YOLO26's training efficiency is substantially better, utilizing far less GPU VRAM than the memory-intensive attention mechanisms of RTDETRv2, allowing for larger batch sizes even on consumer-grade hardware. ```python from ultralytics import YOLO @@ -134,7 +134,7 @@ model.export(format="onnx") ### A Well-Maintained Ecosystem -By utilizing Ultralytics models, developers gain access to an actively maintained framework that integrates natively with modern tracking tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) and [Comet ML](https://docs.ultralytics.com/integrations/comet/). For those who prefer a no-code approach, the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolo26) facilitates cloud training, dataset management, and one-click deployment. +By utilizing Ultralytics models, developers gain access to an actively maintained framework that integrates natively with modern tracking tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) and [Comet ML](https://docs.ultralytics.com/integrations/comet). For those who prefer a no-code approach, the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolo26) facilitates cloud training, dataset management, and one-click deployment. ### Performance Balance @@ -142,4 +142,4 @@ YOLO26 strikes an unparalleled balance between inference speed and accuracy. The ## Other Models in the Ecosystem -While YOLO26 and RTDETRv2 cover the cutting edge of real-time detection, developers maintaining legacy pipelines or exploring different efficiency curves might also consider [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) for established enterprise environments, or explore other architectures like [EfficientDet](https://docs.ultralytics.com/compare/efficientdet-vs-yolov8/). However, for any new initiative, YOLO26 stands as the definitive recommendation. +While YOLO26 and RTDETRv2 cover the cutting edge of real-time detection, developers maintaining legacy pipelines or exploring different efficiency curves might also consider [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) for established enterprise environments, or explore other architectures like [EfficientDet](https://docs.ultralytics.com/compare/efficientdet-vs-yolov8). However, for any new initiative, YOLO26 stands as the definitive recommendation. diff --git a/docs/en/compare/yolo26-vs-yolo11.md b/docs/en/compare/yolo26-vs-yolo11.md index fce4c3423a3..a350d5c092d 100644 --- a/docs/en/compare/yolo26-vs-yolo11.md +++ b/docs/en/compare/yolo26-vs-yolo11.md @@ -23,7 +23,7 @@ Both models were developed by Ultralytics, but they represent different paradigm - **Organization:** [Ultralytics](https://www.ultralytics.com/about) - **Date:** 2026-01-14 - **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO26 Official Documentation](https://docs.ultralytics.com/models/yolo26/) +- **Docs:** [YOLO26 Official Documentation](https://docs.ultralytics.com/models/yolo26) [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -33,21 +33,21 @@ Both models were developed by Ultralytics, but they represent different paradigm - **Organization:** [Ultralytics](https://www.ultralytics.com/about) - **Date:** 2024-09-27 - **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO11 Official Documentation](https://docs.ultralytics.com/models/yolo11/) +- **Docs:** [YOLO11 Official Documentation](https://docs.ultralytics.com/models/yolo11) [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } !!! tip "Other Architectures" - While YOLO26 is our most advanced real-time model, users dealing with highly specialized hardware or massive memory capacities might also explore transformer-based architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) or the breakthrough NMS-free pioneer, [YOLOv10](https://docs.ultralytics.com/models/yolov10/). + While YOLO26 is our most advanced real-time model, users dealing with highly specialized hardware or massive memory capacities might also explore transformer-based architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) or the breakthrough NMS-free pioneer, [YOLOv10](https://docs.ultralytics.com/models/yolov10). ## Architectural Differences and Innovations -The leap from YOLO11 to YOLO26 involves fundamental shifts in both model architecture and the underlying training regimen. While YOLO11 established a robust baseline for [object detection](https://docs.ultralytics.com/tasks/detect/) and multi-task learning, YOLO26 completely overhauls the deployment pipeline for edge computing. +The leap from YOLO11 to YOLO26 involves fundamental shifts in both model architecture and the underlying training regimen. While YOLO11 established a robust baseline for [object detection](https://docs.ultralytics.com/tasks/detect) and multi-task learning, YOLO26 completely overhauls the deployment pipeline for edge computing. ### End-to-End NMS-Free Design -One of the most significant upgrades in YOLO26 is its natively end-to-end architecture. Unlike YOLO11, which relies on [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing to filter overlapping bounding boxes, YOLO26 eliminates this step entirely. This concept, first pioneered in [YOLOv10](https://docs.ultralytics.com/compare/yolov10-vs-yolo26/), dramatically reduces latency variability and simplifies deployment logic across diverse edge devices. +One of the most significant upgrades in YOLO26 is its natively end-to-end architecture. Unlike YOLO11, which relies on [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing to filter overlapping bounding boxes, YOLO26 eliminates this step entirely. This concept, first pioneered in [YOLOv10](https://docs.ultralytics.com/compare/yolov10-vs-yolo26), dramatically reduces latency variability and simplifies deployment logic across diverse edge devices. ### DFL Removal for Edge Efficiency @@ -59,7 +59,7 @@ Training stability and speed are paramount. YOLO26 introduces the **MuSGD Optimi ### ProgLoss and STAL -For researchers working with [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or drone applications, detecting tiny features is a historic challenge. YOLO26 introduces ProgLoss combined with STAL (Scale-Targeted Attention Loss), delivering notable improvements in small-object recognition over YOLO11. +For researchers working with [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or drone applications, detecting tiny features is a historic challenge. YOLO26 introduces ProgLoss combined with STAL (Scale-Targeted Attention Loss), delivering notable improvements in small-object recognition over YOLO11. ## Performance and Metrics Comparison @@ -85,10 +85,10 @@ _Note: The YOLO26 nano (YOLO26n) model showcases a ~31% improvement in CPU speed Both models benefit from the highly maintained Ultralytics ecosystem, offering unparalleled ease of use through a unified Python API. They are not just object detectors; they are multi-task powerhouses. However, YOLO26 incorporates several task-specific advancements: -- **Instance Segmentation:** YOLO26 uses a refined semantic segmentation loss and multi-scale prototyping, generating crisper mask boundaries than YOLO11. Learn more about [segmentation workflows](https://docs.ultralytics.com/tasks/segment/). -- **Pose Estimation:** By integrating Residual Log-Likelihood Estimation (RLE), YOLO26 dramatically improves keypoint accuracy in complex human poses. Discover [pose estimation capabilities](https://docs.ultralytics.com/tasks/pose/). -- **Oriented Bounding Boxes (OBB):** A specialized angle loss function resolves historical boundary discontinuity issues, making YOLO26 exceptionally reliable for detecting rotated objects in satellite feeds. Read about [OBB tasks](https://docs.ultralytics.com/tasks/obb/). -- **Image Classification:** Both models handle high-speed [classification](https://docs.ultralytics.com/tasks/classify/) efficiently, with YOLO26 delivering marginal top-1 accuracy improvements on ImageNet. +- **Instance Segmentation:** YOLO26 uses a refined semantic segmentation loss and multi-scale prototyping, generating crisper mask boundaries than YOLO11. Learn more about [segmentation workflows](https://docs.ultralytics.com/tasks/segment). +- **Pose Estimation:** By integrating Residual Log-Likelihood Estimation (RLE), YOLO26 dramatically improves keypoint accuracy in complex human poses. Discover [pose estimation capabilities](https://docs.ultralytics.com/tasks/pose). +- **Oriented Bounding Boxes (OBB):** A specialized angle loss function resolves historical boundary discontinuity issues, making YOLO26 exceptionally reliable for detecting rotated objects in satellite feeds. Read about [OBB tasks](https://docs.ultralytics.com/tasks/obb). +- **Image Classification:** Both models handle high-speed [classification](https://docs.ultralytics.com/tasks/classify) efficiently, with YOLO26 delivering marginal top-1 accuracy improvements on ImageNet. ## Training and Inference Code Example @@ -139,4 +139,4 @@ While YOLO26 is superior, YOLO11 remains an incredibly capable model. You might Regardless of whether you deploy YOLO11 or YOLO26, utilizing Ultralytics models means tapping into a [well-maintained ecosystem](https://github.com/ultralytics/ultralytics) with frequent updates and vast community support. -For enterprise teams, the [Ultralytics Platform](https://platform.ultralytics.com/) provides an end-to-end solution for [data annotation](https://docs.ultralytics.com/platform/data/annotation/), model training, and seamless cloud deployment. From exporting your trained weights to [CoreML](https://docs.ultralytics.com/integrations/coreml/) or [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), to configuring advanced [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/), the tools provided ensure your AI lifecycle is as streamlined as possible. +For enterprise teams, the [Ultralytics Platform](https://platform.ultralytics.com/) provides an end-to-end solution for [data annotation](https://docs.ultralytics.com/platform/data/annotation), model training, and seamless cloud deployment. From exporting your trained weights to [CoreML](https://docs.ultralytics.com/integrations/coreml) or [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), to configuring advanced [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning), the tools provided ensure your AI lifecycle is as streamlined as possible. diff --git a/docs/en/compare/yolo26-vs-yolov10.md b/docs/en/compare/yolo26-vs-yolov10.md index 1572c941335..a682506b3cf 100644 --- a/docs/en/compare/yolo26-vs-yolov10.md +++ b/docs/en/compare/yolo26-vs-yolov10.md @@ -15,13 +15,13 @@ The landscape of computer vision is constantly evolving, driven by the demand fo ## The Evolution of NMS-Free Architectures -For years, the YOLO (You Only Look Once) family relied heavily on [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) to filter out redundant bounding boxes during post-processing. While effective, NMS introduces inference latency and complicates deployment on edge devices like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or specialized neural processing units (NPUs). +For years, the YOLO (You Only Look Once) family relied heavily on [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) to filter out redundant bounding boxes during post-processing. While effective, NMS introduces inference latency and complicates deployment on edge devices like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or specialized neural processing units (NPUs). The introduction of YOLOv10 represented a paradigm shift by pioneering an end-to-end NMS-free design. Building upon this foundational breakthrough, Ultralytics YOLO26 refined the architecture for production environments, achieving unprecedented efficiency and ease of use across a wider variety of tasks. !!! info "The Post-Processing Bottleneck" - Removing NMS eliminates the dynamic, data-dependent post-processing step that traditionally hindered the optimization of computer vision models on hardware accelerators like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino/). + Removing NMS eliminates the dynamic, data-dependent post-processing step that traditionally hindered the optimization of computer vision models on hardware accelerators like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino). ## YOLOv10: Pioneering NMS-Free Detection @@ -42,7 +42,7 @@ Developed by researchers at Tsinghua University, YOLOv10 introduced a consistent - **Limited Task Support:** Primarily focused on standard object detection, lacking native out-of-the-box support for advanced tasks like segmentation or pose estimation. - **Academic Focus:** The codebase, while robust, leans more toward research rather than streamlined, enterprise-grade production deployment. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## YOLO26: The New Standard for Edge and Cloud @@ -58,8 +58,8 @@ YOLO26 introduces several architectural breakthroughs: - **DFL Removal:** Distribution Focal Loss has been completely removed. This dramatically simplifies the model export process and improves compatibility with edge and low-power devices. - **Up to 43% Faster CPU Inference:** Thanks to DFL removal and structural optimizations, YOLO26 is significantly faster on CPUs, making it ideal for IoT and mobile deployments. - **MuSGD Optimizer:** Inspired by Large Language Model (LLM) training techniques (such as Moonshot AI's Kimi K2), YOLO26 utilizes a hybrid of SGD and Muon. This brings unparalleled training stability and faster convergence to computer vision. -- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, which is critical for [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and drone-based [security monitoring](https://www.ultralytics.com/blog/real-time-security-monitoring-with-ai-and-ultralytics-yolo11). -- **Task-Specific Improvements:** YOLO26 isn't just a detector. It features Semantic Segmentation loss and multi-scale proto for [Segmentation](https://docs.ultralytics.com/tasks/segment/), Residual Log-Likelihood Estimation (RLE) for [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), and specialized angle loss for [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, which is critical for [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and drone-based [security monitoring](https://www.ultralytics.com/blog/real-time-security-monitoring-with-ai-and-ultralytics-yolo11). +- **Task-Specific Improvements:** YOLO26 isn't just a detector. It features Semantic Segmentation loss and multi-scale proto for [Segmentation](https://docs.ultralytics.com/tasks/segment), Residual Log-Likelihood Estimation (RLE) for [Pose Estimation](https://docs.ultralytics.com/tasks/pose), and specialized angle loss for [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -84,11 +84,11 @@ The following table compares the COCO detection performance of YOLO26 and YOLOv1 ### The Ultralytics Advantage: Training and Memory Efficiency -When deploying models into production, memory requirements and training efficiency are just as crucial as inference speed. Ultralytics models, particularly YOLO26, are highly optimized to reduce CUDA memory usage during training. This allows developers to use larger [batch sizes](https://www.ultralytics.com/glossary/batch-size) on consumer-grade GPUs, drastically cutting down training time and computational costs. Conversely, complex architectures or heavy transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) often require expensive, high-end hardware to train effectively. +When deploying models into production, memory requirements and training efficiency are just as crucial as inference speed. Ultralytics models, particularly YOLO26, are highly optimized to reduce CUDA memory usage during training. This allows developers to use larger [batch sizes](https://www.ultralytics.com/glossary/batch-size) on consumer-grade GPUs, drastically cutting down training time and computational costs. Conversely, complex architectures or heavy transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) often require expensive, high-end hardware to train effectively. !!! tip "Continuous Integration and Ecosystem" - One of the greatest benefits of choosing YOLO26 is its integration with the well-maintained Ultralytics ecosystem. From [data annotation](https://docs.ultralytics.com/platform/data/annotation/) to [experiment tracking](https://docs.ultralytics.com/integrations/weights-biases/), the platform provides everything a machine learning engineer needs under one unified roof. + One of the greatest benefits of choosing YOLO26 is its integration with the well-maintained Ultralytics ecosystem. From [data annotation](https://docs.ultralytics.com/platform/data/annotation) to [experiment tracking](https://docs.ultralytics.com/integrations/weights-biases), the platform provides everything a machine learning engineer needs under one unified roof. ## Practical Implementation: Code Example @@ -132,7 +132,7 @@ YOLO26 is a strong choice for: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ### When to Choose YOLOv10 @@ -146,4 +146,4 @@ YOLOv10 is recommended for: While YOLOv10 made significant contributions to the academic community by introducing the NMS-free paradigm, **YOLO26** elevates this technology to enterprise-grade readiness. With its remarkable 43% boost in CPU speed, the innovative MuSGD optimizer, and unmatched versatility across vision tasks, YOLO26 stands out as the ultimate choice for both edge computing and large-scale cloud deployments. -For teams prioritizing an active community, comprehensive [documentation](https://docs.ultralytics.com/), and a frictionless developer experience, the Ultralytics ecosystem is unparalleled. If you are exploring models for specialized scenarios, you may also want to investigate [YOLO-World](https://docs.ultralytics.com/models/yolo-world/) for zero-shot open-vocabulary detection. However, for the vast majority of real-world use cases, **YOLO26** is the definitive recommendation. +For teams prioritizing an active community, comprehensive [documentation](https://docs.ultralytics.com/), and a frictionless developer experience, the Ultralytics ecosystem is unparalleled. If you are exploring models for specialized scenarios, you may also want to investigate [YOLO-World](https://docs.ultralytics.com/models/yolo-world) for zero-shot open-vocabulary detection. However, for the vast majority of real-world use cases, **YOLO26** is the definitive recommendation. diff --git a/docs/en/compare/yolo26-vs-yolov5.md b/docs/en/compare/yolo26-vs-yolov5.md index e3409e2a629..7a4d5609556 100644 --- a/docs/en/compare/yolo26-vs-yolov5.md +++ b/docs/en/compare/yolo26-vs-yolov5.md @@ -23,7 +23,7 @@ Before diving into architectural nuances, let's establish the foundational detai - Organization: [Ultralytics](https://www.ultralytics.com/) - Date: 2026-01-14 - GitHub: [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- Docs: [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26/) +- Docs: [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26) [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -33,13 +33,13 @@ Before diving into architectural nuances, let's establish the foundational detai - Organization: [Ultralytics](https://www.ultralytics.com/) - Date: 2020-06-26 - GitHub: [https://github.com/ultralytics/yolov5](https://github.com/ultralytics/yolov5) -- Docs: [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5/) +- Docs: [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5) [Learn more about YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5){ .md-button } !!! tip "Exploring Other Options" - While this guide focuses on YOLO26 and YOLOv5, developers migrating legacy systems might also be interested in comparing [YOLO11](https://docs.ultralytics.com/models/yolo11/) or the pioneering NMS-free architecture of [YOLOv10](https://docs.ultralytics.com/models/yolov10/). Both offer excellent stepping stones for specific deployment environments. + While this guide focuses on YOLO26 and YOLOv5, developers migrating legacy systems might also be interested in comparing [YOLO11](https://docs.ultralytics.com/models/yolo11) or the pioneering NMS-free architecture of [YOLOv10](https://docs.ultralytics.com/models/yolov10). Both offer excellent stepping stones for specific deployment environments. ## Architectural Innovations @@ -89,7 +89,7 @@ predictions = model("https://ultralytics.com/images/bus.jpg") predictions[0].show() ``` -This simple script allows developers to rapidly iterate on [custom datasets](https://docs.ultralytics.com/datasets/), seamlessly moving from data ingestion to a production-ready model. +This simple script allows developers to rapidly iterate on [custom datasets](https://docs.ultralytics.com/datasets), seamlessly moving from data ingestion to a production-ready model. !!! note "Deployment Made Easy" @@ -103,7 +103,7 @@ YOLOv5 remains a reliable workhorse for legacy systems. If you have an existing ### When to Use YOLO26 -YOLO26 is the definitive choice for modern computer vision projects. Its versatility far outstrips its predecessor. While YOLOv5 primarily focuses on detection (with later segmentation additions), YOLO26 offers deep, native support for [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +YOLO26 is the definitive choice for modern computer vision projects. Its versatility far outstrips its predecessor. While YOLOv5 primarily focuses on detection (with later segmentation additions), YOLO26 offers deep, native support for [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Image Classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). YOLO26 introduces **Task-Specific Improvements**, such as a specialized semantic segmentation loss, Residual Log-Likelihood Estimation (RLE) for ultra-precise pose keypoints, and advanced angle loss for OBB to solve tricky boundary issues. diff --git a/docs/en/compare/yolo26-vs-yolov6.md b/docs/en/compare/yolo26-vs-yolov6.md index 170392d0746..a091def8913 100644 --- a/docs/en/compare/yolo26-vs-yolov6.md +++ b/docs/en/compare/yolo26-vs-yolov6.md @@ -23,7 +23,7 @@ Before diving into performance metrics, it is helpful to understand the backgrou - Organization: [Ultralytics](https://www.ultralytics.com/) - Date: 2026-01-14 - GitHub: [Ultralytics GitHub Repository](https://github.com/ultralytics/ultralytics) -- Docs: [YOLO26 Official Documentation](https://docs.ultralytics.com/models/yolo26/) +- Docs: [YOLO26 Official Documentation](https://docs.ultralytics.com/models/yolo26) [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -34,23 +34,23 @@ Before diving into performance metrics, it is helpful to understand the backgrou - Date: 2023-01-13 - Arxiv: [YOLOv6 v3.0 Paper](https://arxiv.org/abs/2301.05586) - GitHub: [YOLOv6 GitHub Repository](https://github.com/meituan/YOLOv6) -- Docs: [YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6/) +- Docs: [YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6) -[Learn more about YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Architectural Innovations and Differences -Both models are designed for high-speed [object detection](https://docs.ultralytics.com/tasks/detect/), but they take vastly different approaches to achieve their performance. +Both models are designed for high-speed [object detection](https://docs.ultralytics.com/tasks/detect), but they take vastly different approaches to achieve their performance. ### Ultralytics YOLO26: The Edge-First Native End-to-End Model -Released in early 2026, YOLO26 represents a massive leap forward in model efficiency. The most significant architectural upgrade is its natively **End-to-End NMS-Free Design**. By eliminating the traditional [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing step—a concept successfully pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/)—YOLO26 drastically reduces latency variability, making it highly predictable for real-time edge deployments. +Released in early 2026, YOLO26 represents a massive leap forward in model efficiency. The most significant architectural upgrade is its natively **End-to-End NMS-Free Design**. By eliminating the traditional [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing step—a concept successfully pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10)—YOLO26 drastically reduces latency variability, making it highly predictable for real-time edge deployments. Additionally, YOLO26 features **DFL Removal**. By stripping out the Distribution Focal Loss, the model simplifies its export process and significantly enhances compatibility with low-power [edge computing](https://www.ultralytics.com/glossary/edge-computing) devices. This results in up to **43% Faster CPU Inference**, making YOLO26 an absolute powerhouse for environments without dedicated [graphics processing units (GPUs)](https://www.ultralytics.com/glossary/gpu-graphics-processing-unit) like Raspberry Pi or mobile devices. ### YOLOv6-3.0: The Industrial Specialist -Developed by the vision team at Meituan, YOLOv6-3.0 is a highly capable, industrial-grade CNN heavily optimized for [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) deployment on NVIDIA hardware. It relies heavily on self-distillation techniques and hardware-aware neural architecture design. While incredibly fast on heavy T4 or A100 GPUs, it relies on traditional NMS post-processing, which can introduce bottlenecks in constrained hardware environments. +Developed by the vision team at Meituan, YOLOv6-3.0 is a highly capable, industrial-grade CNN heavily optimized for [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) deployment on NVIDIA hardware. It relies heavily on self-distillation techniques and hardware-aware neural architecture design. While incredibly fast on heavy T4 or A100 GPUs, it relies on traditional NMS post-processing, which can introduce bottlenecks in constrained hardware environments. ## Performance Balance and Benchmarks @@ -91,7 +91,7 @@ In contrast, YOLOv6 requires cloning the research repository, managing dependenc ### Code Example: Getting Started with YOLO26 -Training and running inference with Ultralytics models is brilliantly simple. The robust [Python API](https://docs.ultralytics.com/usage/python/) handles all the heavy lifting: +Training and running inference with Ultralytics models is brilliantly simple. The robust [Python API](https://docs.ultralytics.com/usage/python) handles all the heavy lifting: ```python from ultralytics import YOLO @@ -111,7 +111,7 @@ model.export(format="onnx") ## Unmatched Versatility Across Vision Tasks -While YOLOv6-3.0 is strictly a bounding-box object detector, YOLO26 boasts incredible versatility. Using the exact same simple API, developers can perform [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. +While YOLOv6-3.0 is strictly a bounding-box object detector, YOLO26 boasts incredible versatility. Using the exact same simple API, developers can perform [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. YOLO26 includes task-specific improvements across the board, such as semantic segmentation loss for pixel-perfect masking, Residual Log-Likelihood Estimation (RLE) for hyper-accurate keypoints, and specialized angle loss to resolve OBB boundary issues. @@ -119,7 +119,7 @@ YOLO26 includes task-specific improvements across the board, such as semantic se ### When to use YOLO26 -YOLO26 is the undisputed champion for edge devices, [Internet of Things (IoT)](https://www.ultralytics.com/blog/industrial-iot-iiot-internet-of-things-explained), and robotics. Its 43% faster CPU inference and NMS-free architecture make it perfect for real-time [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/) running on standard CPUs or low-power ARM chips. Its superior small object detection (thanks to ProgLoss + STAL) makes it the ideal candidate for aerial [wildlife detection](https://www.ultralytics.com/blog/wildlife-detection-for-your-backyard-powered-by-vision-ai) and satellite imagery analysis. +YOLO26 is the undisputed champion for edge devices, [Internet of Things (IoT)](https://www.ultralytics.com/blog/industrial-iot-iiot-internet-of-things-explained), and robotics. Its 43% faster CPU inference and NMS-free architecture make it perfect for real-time [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system) running on standard CPUs or low-power ARM chips. Its superior small object detection (thanks to ProgLoss + STAL) makes it the ideal candidate for aerial [wildlife detection](https://www.ultralytics.com/blog/wildlife-detection-for-your-backyard-powered-by-vision-ai) and satellite imagery analysis. ### When to use YOLOv6-3.0 @@ -127,7 +127,7 @@ YOLOv6-3.0 shines in tightly controlled industrial environments where servers ar ## Exploring Other Models -If you are exploring the broader landscape of computer vision, you may also be interested in other models supported by the Ultralytics ecosystem. For instance, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a fantastic general-purpose model with massive community backing. If you are specifically interested in transformer architectures, the [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) model offers robust attention-based performance, though it requires significantly more training memory than YOLO26. For zero-shot capabilities without training, [YOLO-World](https://docs.ultralytics.com/models/yolo-world/) provides promptable open-vocabulary detection out of the box. +If you are exploring the broader landscape of computer vision, you may also be interested in other models supported by the Ultralytics ecosystem. For instance, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a fantastic general-purpose model with massive community backing. If you are specifically interested in transformer architectures, the [RT-DETR](https://docs.ultralytics.com/models/rtdetr) model offers robust attention-based performance, though it requires significantly more training memory than YOLO26. For zero-shot capabilities without training, [YOLO-World](https://docs.ultralytics.com/models/yolo-world) provides promptable open-vocabulary detection out of the box. ## Summary diff --git a/docs/en/compare/yolo26-vs-yolov7.md b/docs/en/compare/yolo26-vs-yolov7.md index 4e6daaca8d4..68cdaed7289 100644 --- a/docs/en/compare/yolo26-vs-yolov7.md +++ b/docs/en/compare/yolo26-vs-yolov7.md @@ -25,7 +25,7 @@ Before examining the performance data, it is important to understand the origins **Organization:** [Ultralytics](https://www.ultralytics.com) **Date:** 2026-01-14 **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -**Docs:** [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26/) +**Docs:** [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26) [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -37,11 +37,11 @@ Before examining the performance data, it is important to understand the origins **Arxiv:** [YOLOv7 Paper](https://arxiv.org/abs/2207.02696) **GitHub:** [YOLOv7 Repository](https://github.com/WongKinYiu/yolov7) -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } !!! note "Alternative Models to Consider" - If you are exploring the broader ecosystem, you might also be interested in [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) for highly balanced multi-task deployments, or the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) for sequence-based detection. Note that older models like [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) and [YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5) remain fully supported on the Ultralytics Platform for legacy integration. + If you are exploring the broader ecosystem, you might also be interested in [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) for highly balanced multi-task deployments, or the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr) for sequence-based detection. Note that older models like [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) and [YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5) remain fully supported on the Ultralytics Platform for legacy integration. ## Architectural Deep Dive @@ -49,7 +49,7 @@ The architectural philosophies behind YOLO26 and YOLOv7 diverge significantly, r ### YOLO26: The Edge-First Paradigm -Released in 2026, YOLO26 fundamentally rethinks the deployment pipeline. Its most significant breakthrough is the **End-to-End NMS-Free Design**. By eliminating [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing, YOLO26 drastically reduces latency variability, a concept that was first successfully piloted in [YOLOv10](https://docs.ultralytics.com/models/yolov10/). This ensures consistent frame rates even in densely populated scenes, which is critical for autonomous robotics and traffic monitoring. +Released in 2026, YOLO26 fundamentally rethinks the deployment pipeline. Its most significant breakthrough is the **End-to-End NMS-Free Design**. By eliminating [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing, YOLO26 drastically reduces latency variability, a concept that was first successfully piloted in [YOLOv10](https://docs.ultralytics.com/models/yolov10). This ensures consistent frame rates even in densely populated scenes, which is critical for autonomous robotics and traffic monitoring. Furthermore, YOLO26 completely removes Distribution Focal Loss (DFL). This **DFL Removal** simplifies the export process to formats like [ONNX](https://onnx.ai/) and [Apple CoreML](https://developer.apple.com/machine-learning/core-ml/), achieving up to **43% faster CPU inference**. @@ -80,11 +80,11 @@ _Note: YOLO26x outperforms YOLOv7x in mAP by an impressive margin (57.5 vs 53.1) ## The Ultralytics Ecosystem Advantage -A primary reason developers consistently choose YOLO26 is its deep integration into the [Ultralytics Platform](https://docs.ultralytics.com/platform/). Unlike the standalone scripts required for older architectures, Ultralytics provides a seamless, unified workflow. +A primary reason developers consistently choose YOLO26 is its deep integration into the [Ultralytics Platform](https://docs.ultralytics.com/platform). Unlike the standalone scripts required for older architectures, Ultralytics provides a seamless, unified workflow. 1. **Ease of Use:** The Python API allows users to load, train, and deploy models in just a few lines of code. Exporting to mobile formats like [TensorFlow Lite](https://ai.google.dev/edge/litert) requires merely changing a single argument. 2. **Memory Requirements:** Ultralytics models are meticulously engineered for training efficiency. They require significantly less CUDA memory compared to heavy vision transformer models, allowing researchers to run larger batch sizes on consumer hardware. -3. **Versatility:** While YOLOv7 requires entirely different repositories for different tasks, YOLO26 natively supports [Image Classification](https://docs.ultralytics.com/tasks/classify/), [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection from a single, cohesive library. It even includes task-specific loss functions, such as Residual Log-Likelihood Estimation (RLE) for human pose pipelines. +3. **Versatility:** While YOLOv7 requires entirely different repositories for different tasks, YOLO26 natively supports [Image Classification](https://docs.ultralytics.com/tasks/classify), [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection from a single, cohesive library. It even includes task-specific loss functions, such as Residual Log-Likelihood Estimation (RLE) for human pose pipelines. 4. **Active Development:** The Ultralytics open-source community provides frequent updates, ensuring rapid resolution of edge cases and continuous compatibility with the latest [PyTorch](https://pytorch.org/) releases. !!! tip "Streamlined Exporting" diff --git a/docs/en/compare/yolo26-vs-yolov8.md b/docs/en/compare/yolo26-vs-yolov8.md index a45c4ac0e1c..2389498cc19 100644 --- a/docs/en/compare/yolo26-vs-yolov8.md +++ b/docs/en/compare/yolo26-vs-yolov8.md @@ -8,7 +8,7 @@ keywords: YOLO26, YOLOv8, YOLO comparison, object detection, NMS-free, end-to-en The evolution of computer vision has been defined by the pursuit of real-time performance without sacrificing accuracy. As developers and researchers navigate the landscape of modern [machine learning](https://en.wikipedia.org/wiki/Machine_learning), choosing the right model architecture is critical. This comprehensive technical comparison explores the generational leap from **[Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8)**, a wildly popular architecture that redefined the standard in 2023, to the cutting-edge **[Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26)**, released in January 2026. -By delving into their architectures, performance metrics, and training methodologies, we highlight why upgrading to the latest innovations provides distinct advantages for [object detection](https://docs.ultralytics.com/tasks/detect/), segmentation, and beyond. +By delving into their architectures, performance metrics, and training methodologies, we highlight why upgrading to the latest innovations provides distinct advantages for [object detection](https://docs.ultralytics.com/tasks/detect), segmentation, and beyond. @@ -24,7 +24,7 @@ Authors: Glenn Jocher and Jing Qiu Organization: [Ultralytics](https://www.ultralytics.com/) Date: 2026-01-14 GitHub: [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -Docs: [https://docs.ultralytics.com/models/yolo26/](https://docs.ultralytics.com/models/yolo26/) +Docs: [https://docs.ultralytics.com/models/yolo26/](https://docs.ultralytics.com/models/yolo26) [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -33,7 +33,7 @@ Authors: Glenn Jocher, Ayush Chaurasia, and Jing Qiu Organization: [Ultralytics](https://www.ultralytics.com/) Date: 2023-01-10 GitHub: [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -Docs: [https://docs.ultralytics.com/models/yolov8/](https://docs.ultralytics.com/models/yolov8/) +Docs: [https://docs.ultralytics.com/models/yolov8/](https://docs.ultralytics.com/models/yolov8) [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -45,14 +45,14 @@ The transition from YOLOv8 to YOLO26 introduces significant paradigm shifts in h YOLO26 was engineered from the ground up to eliminate deployment bottlenecks and maximize inference speed on constrained hardware. -- **End-to-End NMS-Free Design:** Building on concepts first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively employs an end-to-end architecture. By completely eliminating the need for Non-Maximum Suppression (NMS) post-processing, latency variance is virtually eradicated. This simplifies deployment logic for applications requiring strict real-time guarantees. +- **End-to-End NMS-Free Design:** Building on concepts first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively employs an end-to-end architecture. By completely eliminating the need for Non-Maximum Suppression (NMS) post-processing, latency variance is virtually eradicated. This simplifies deployment logic for applications requiring strict real-time guarantees. - **DFL Removal:** The removal of Distribution Focal Loss (DFL) drastically simplifies the output head. This architectural choice enables significantly better compatibility with low-power edge devices and simpler exports to formats like [ONNX](https://onnx.ai/) and [CoreML](https://developer.apple.com/machine-learning/core-ml/). - **MuSGD Optimizer:** Inspired by the training stability seen in Large Language Models (LLMs) like Moonshot AI's Kimi K2, YOLO26 utilizes the MuSGD optimizer—a hybrid of Stochastic Gradient Descent and Muon. This brings LLM-scale training innovations into computer vision, yielding faster convergence and highly stable training runs. -- **ProgLoss + STAL:** To combat the notoriously difficult problem of recognizing tiny subjects, YOLO26 implements Progressive Loss (ProgLoss) combined with Scale-Tolerant Anchor Loss (STAL). This provides critical improvements for [small object detection](https://docs.ultralytics.com/guides/vision-eye/), making it ideal for drone applications. +- **ProgLoss + STAL:** To combat the notoriously difficult problem of recognizing tiny subjects, YOLO26 implements Progressive Loss (ProgLoss) combined with Scale-Tolerant Anchor Loss (STAL). This provides critical improvements for [small object detection](https://docs.ultralytics.com/guides/vision-eye), making it ideal for drone applications. !!! info "Task-Specific Refinements" - YOLO26 also brings targeted upgrades across multiple computer vision domains. It utilizes a Semantic Segmentation loss and multi-scale proto for better [instance segmentation](https://docs.ultralytics.com/tasks/segment/), Residual Log-Likelihood Estimation (RLE) for highly accurate [pose estimation](https://docs.ultralytics.com/tasks/pose/), and specialized angle loss algorithms to resolve boundary issues in [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). + YOLO26 also brings targeted upgrades across multiple computer vision domains. It utilizes a Semantic Segmentation loss and multi-scale proto for better [instance segmentation](https://docs.ultralytics.com/tasks/segment), Residual Log-Likelihood Estimation (RLE) for highly accurate [pose estimation](https://docs.ultralytics.com/tasks/pose), and specialized angle loss algorithms to resolve boundary issues in [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). ### YOLOv8: The Highly Versatile Workhorse @@ -60,7 +60,7 @@ When released in 2023, YOLOv8 set a new benchmark by fully transitioning to an a - **C2f Module:** It replaced the older C3 module with the C2f block, allowing for better gradient flow across the network backbone. - **Decoupled Head:** YOLOv8 features a decoupled head where classification and bounding box regression are computed independently, significantly boosting the mean Average Precision (mAP). -- **Task Versatility:** It was one of the first models to provide a truly unified API for [image classification](https://docs.ultralytics.com/tasks/classify/), detection, segmentation, and pose tasks out of the box. +- **Task Versatility:** It was one of the first models to provide a truly unified API for [image classification](https://docs.ultralytics.com/tasks/classify), detection, segmentation, and pose tasks out of the box. ## Performance Metrics and Resource Requirements @@ -93,8 +93,8 @@ Furthermore, YOLO26 models feature lower parameter counts and FLOPs for their re A major consideration when selecting an AI model is the surrounding infrastructure. Both YOLO26 and YOLOv8 benefit immensely from the unified [Ultralytics Platform](https://platform.ultralytics.com), providing an unparalleled developer experience. 1. **Ease of Use:** The "zero-to-hero" philosophy ensures developers can load, train, and export models in minimal code. The Python API remains consistent across model generations. -2. **Training Efficiency:** Ultralytics YOLO models require exceptionally lower CUDA memory during training runs compared to transformer models (like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/)). This permits the use of larger batch sizes on consumer hardware, democratizing AI research. -3. **Well-Maintained Ecosystem:** Backed by continuous updates, rigorous CI/CD pipelines, and deep integrations with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) and [TensorRT](https://developer.nvidia.com/tensorrt), the Ultralytics repository is robust and production-ready. +2. **Training Efficiency:** Ultralytics YOLO models require exceptionally lower CUDA memory during training runs compared to transformer models (like [RT-DETR](https://docs.ultralytics.com/models/rtdetr)). This permits the use of larger batch sizes on consumer hardware, democratizing AI research. +3. **Well-Maintained Ecosystem:** Backed by continuous updates, rigorous CI/CD pipelines, and deep integrations with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) and [TensorRT](https://developer.nvidia.com/tensorrt), the Ultralytics repository is robust and production-ready. 4. **Unmatched Versatility:** Ultralytics models are not one-trick ponies; a single import handles diverse datasets, augmenting workflows for complex systems that require simultaneous tracking, classification, and segmentation. !!! tip "Streamlined Upgrades" @@ -107,11 +107,11 @@ Choosing between these models often comes down to your deployment constraints, t ### Edge Computing and IoT Networks -For edge environments—such as [Raspberry Pi deployments](https://docs.ultralytics.com/guides/raspberry-pi/) or localized factory floor sensors—**YOLO26** is the undisputed champion. Its natively optimized CPU speed and NMS-free structure mean smart cameras can process high-framerate video for [parking management](https://docs.ultralytics.com/guides/parking-management/) without dropping frames due to post-processing bottlenecks. +For edge environments—such as [Raspberry Pi deployments](https://docs.ultralytics.com/guides/raspberry-pi) or localized factory floor sensors—**YOLO26** is the undisputed champion. Its natively optimized CPU speed and NMS-free structure mean smart cameras can process high-framerate video for [parking management](https://docs.ultralytics.com/guides/parking-management) without dropping frames due to post-processing bottlenecks. ### High-Altitude and Aerial Imagery -In [agricultural monitoring](https://www.ultralytics.com/solutions/ai-in-agriculture) or infrastructure inspection via drones, small object detection is paramount. The **ProgLoss + STAL** implementation in **YOLO26** allows it to consistently detect tiny pests or micro-fractures in pipelines that older architectures like YOLOv8 might miss, offering superior recall and precision on datasets like [VisDrone](https://docs.ultralytics.com/datasets/detect/visdrone/). +In [agricultural monitoring](https://www.ultralytics.com/solutions/ai-in-agriculture) or infrastructure inspection via drones, small object detection is paramount. The **ProgLoss + STAL** implementation in **YOLO26** allows it to consistently detect tiny pests or micro-fractures in pipelines that older architectures like YOLOv8 might miss, offering superior recall and precision on datasets like [VisDrone](https://docs.ultralytics.com/datasets/detect/visdrone). ### Legacy GPU Systems @@ -127,13 +127,13 @@ YOLO26 is a strong choice for: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ### When to Choose YOLOv8 YOLOv8 is recommended for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. @@ -168,6 +168,6 @@ predictions[0].show() While YOLO26 represents the current state-of-the-art, developers building diverse applications might also explore: - **[YOLO11](https://platform.ultralytics.com/ultralytics/yolo11)**: The immediate predecessor to YOLO26, offering exceptional refinement over YOLOv8 and still heavily utilized in cutting-edge production systems. -- **[RT-DETR](https://docs.ultralytics.com/models/rtdetr/)**: Baidu's Real-Time DEtection TRansformer. It is an excellent choice for researchers exploring the attention mechanism in vision tasks, though it requires significantly more CUDA memory to train compared to standard Ultralytics YOLO models. +- **[RT-DETR](https://docs.ultralytics.com/models/rtdetr)**: Baidu's Real-Time DEtection TRansformer. It is an excellent choice for researchers exploring the attention mechanism in vision tasks, though it requires significantly more CUDA memory to train compared to standard Ultralytics YOLO models. For a comprehensive suite of cloud training, dataset labeling, and immediate deployment, explore the [Ultralytics Platform](https://platform.ultralytics.com/) today. diff --git a/docs/en/compare/yolo26-vs-yolov9.md b/docs/en/compare/yolo26-vs-yolov9.md index 1c4365dc462..1e537137f11 100644 --- a/docs/en/compare/yolo26-vs-yolov9.md +++ b/docs/en/compare/yolo26-vs-yolov9.md @@ -23,7 +23,7 @@ Released in early 2026, [Ultralytics YOLO26](https://platform.ultralytics.com/ul - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2026-01-14 - **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26/) +- **Docs:** [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26) ### Architecture and Innovations @@ -35,7 +35,7 @@ Additionally, the removal of Distribution Focal Loss (DFL) streamlines the expor Thanks to architectural simplifications and the removal of DFL, YOLO26 achieves up to **43% faster CPU inference**, making it the ideal choice for resource-constrained edge devices like the [Raspberry Pi](https://www.raspberrypi.org/) or [NVIDIA Jetson Nano](https://developer.nvidia.com/embedded/jetson-nano-developer-kit). -For detecting highly challenging items in scenes like [drone aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/), YOLO26 utilizes the updated **ProgLoss + STAL** loss functions. These provide notable improvements in small-object recognition recall. Furthermore, it boasts task-specific enhancements, including multi-scale proto for [instance segmentation](https://docs.ultralytics.com/tasks/segment/), Residual Log-Likelihood Estimation (RLE) for [pose estimation](https://docs.ultralytics.com/tasks/pose/), and specialized angle loss for detecting [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +For detecting highly challenging items in scenes like [drone aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone), YOLO26 utilizes the updated **ProgLoss + STAL** loss functions. These provide notable improvements in small-object recognition recall. Furthermore, it boasts task-specific enhancements, including multi-scale proto for [instance segmentation](https://docs.ultralytics.com/tasks/segment), Residual Log-Likelihood Estimation (RLE) for [pose estimation](https://docs.ultralytics.com/tasks/pose), and specialized angle loss for detecting [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -50,7 +50,7 @@ Introduced in early 2024, YOLOv9 brought theoretical advancements to the way neu - **Date:** 2024-02-21 - **Arxiv:** [YOLOv9 Paper](https://arxiv.org/abs/2402.13616) - **GitHub:** [YOLOv9 Repository](https://github.com/WongKinYiu/yolov9) -- **Docs:** [YOLOv9 Documentation](https://docs.ultralytics.com/models/yolov9/) +- **Docs:** [YOLOv9 Documentation](https://docs.ultralytics.com/models/yolov9) ### Architecture and Strengths @@ -58,13 +58,13 @@ YOLOv9 is built around the concept of Programmable Gradient Information (PGI) an ### Limitations -Despite its excellent parameter efficiency, YOLOv9 relies heavily on traditional NMS for bounding box post-processing, which can create computational bottlenecks during inference on edge devices. Furthermore, the official repository is largely focused on object detection, requiring significant custom engineering to adapt it for specialized tasks like [tracking](https://docs.ultralytics.com/modes/track/) or pose estimation. +Despite its excellent parameter efficiency, YOLOv9 relies heavily on traditional NMS for bounding box post-processing, which can create computational bottlenecks during inference on edge devices. Furthermore, the official repository is largely focused on object detection, requiring significant custom engineering to adapt it for specialized tasks like [tracking](https://docs.ultralytics.com/modes/track) or pose estimation. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ## Performance Comparison -When evaluating these models for real-world deployment, balancing accuracy (mAP), inference speed, and memory usage is critical. Ultralytics models are renowned for their low memory requirements during both training and inference, requiring far less [CUDA memory](https://developer.nvidia.com/cuda) than transformer-based alternatives like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +When evaluating these models for real-world deployment, balancing accuracy (mAP), inference speed, and memory usage is critical. Ultralytics models are renowned for their low memory requirements during both training and inference, requiring far less [CUDA memory](https://developer.nvidia.com/cuda) than transformer-based alternatives like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). Below is a direct comparison of YOLO26 and YOLOv9 performance on the [COCO dataset](https://cocodataset.org/). Best values in each column are highlighted in **bold**. @@ -94,7 +94,7 @@ YOLO26 is a strong choice for: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ### When to Choose YOLOv9 @@ -110,7 +110,7 @@ Choosing a model involves more than just reading an accuracy benchmark; the surr ### Ease of Use and Ecosystem -The [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) offers a seamless "zero-to-hero" experience. Instead of cloning complex repositories or manually configuring distributed training scripts, developers can install the package via `pip` and start training immediately. The actively maintained [Ultralytics ecosystem](https://platform.ultralytics.com/deploy/inference/) guarantees frequent updates, automated integrations with ML platforms like [Weights & Biases](https://wandb.ai/site), and extensive documentation. +The [Ultralytics Python API](https://docs.ultralytics.com/usage/python) offers a seamless "zero-to-hero" experience. Instead of cloning complex repositories or manually configuring distributed training scripts, developers can install the package via `pip` and start training immediately. The actively maintained [Ultralytics ecosystem](https://platform.ultralytics.com/deploy/inference/) guarantees frequent updates, automated integrations with ML platforms like [Weights & Biases](https://wandb.ai/site), and extensive documentation. !!! note "Other Ultralytics Models" @@ -118,7 +118,7 @@ The [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) offers ### Versatility Across Vision Tasks -While YOLOv9 is primarily a detection engine, YOLO26 is a general-purpose vision tool. Using a single unified syntax, you can easily pivot from object detection to pixel-perfect [image segmentation](https://docs.ultralytics.com/tasks/segment/) or whole-image [classification](https://docs.ultralytics.com/tasks/classify/). This versatility reduces the technical debt of maintaining multiple disjointed codebases for different computer vision features. +While YOLOv9 is primarily a detection engine, YOLO26 is a general-purpose vision tool. Using a single unified syntax, you can easily pivot from object detection to pixel-perfect [image segmentation](https://docs.ultralytics.com/tasks/segment) or whole-image [classification](https://docs.ultralytics.com/tasks/classify). This versatility reduces the technical debt of maintaining multiple disjointed codebases for different computer vision features. ### Efficient Training and Deployment diff --git a/docs/en/compare/yolo26-vs-yolox.md b/docs/en/compare/yolo26-vs-yolox.md index 8becd9786ff..d63d81a3f45 100644 --- a/docs/en/compare/yolo26-vs-yolox.md +++ b/docs/en/compare/yolo26-vs-yolox.md @@ -23,7 +23,7 @@ Understanding the origins and core philosophies of these models is essential for - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2026-01-14 - **GitHub:** [Ultralytics GitHub Repository](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO26 Official Documentation](https://docs.ultralytics.com/models/yolo26/) +- **Docs:** [YOLO26 Official Documentation](https://docs.ultralytics.com/models/yolo26) [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -46,7 +46,7 @@ Understanding the origins and core philosophies of these models is essential for The differences between YOLO26 and YOLOX highlight five years of relentless innovation in deep learning design. -While YOLOX championed the anchor-free approach, it still relied heavily on traditional Non-Maximum Suppression (NMS) to filter redundant bounding boxes. YOLO26 introduces an **End-to-End NMS-Free Design**. This breakthrough, first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), completely eliminates NMS post-processing, resulting in faster and simpler deployment pipelines with significantly lower latency variance. +While YOLOX championed the anchor-free approach, it still relied heavily on traditional Non-Maximum Suppression (NMS) to filter redundant bounding boxes. YOLO26 introduces an **End-to-End NMS-Free Design**. This breakthrough, first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), completely eliminates NMS post-processing, resulting in faster and simpler deployment pipelines with significantly lower latency variance. Furthermore, YOLO26 features **DFL Removal**. By removing the Distribution Focal Loss, the model's export process is drastically simplified, ensuring exceptional compatibility with edge devices and low-power hardware. When combined with the model's architectural optimizations, YOLO26 achieves up to **43% faster CPU inference** compared to its predecessors, making it a powerhouse for environments lacking dedicated GPUs. @@ -54,11 +54,11 @@ Training stability is another critical differentiator. YOLO26 utilizes the novel !!! note "Advanced Loss Functions" - YOLO26 utilizes **ProgLoss + STAL**, specialized loss functions that yield notable improvements in small-object recognition. This is critical for complex tasks like processing [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and analyzing dense environments. + YOLO26 utilizes **ProgLoss + STAL**, specialized loss functions that yield notable improvements in small-object recognition. This is critical for complex tasks like processing [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and analyzing dense environments. ## Performance and Benchmarks -When comparing these models head-to-head on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/), YOLO26's superiority in both accuracy and efficiency becomes clear. Ultralytics models consistently offer lower memory requirements during training and faster inference speeds. +When comparing these models head-to-head on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco), YOLO26's superiority in both accuracy and efficiency becomes clear. Ultralytics models consistently offer lower memory requirements during training and faster inference speeds. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | --------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -81,7 +81,7 @@ _Note: The YOLO26x model achieves an impressive 57.5 mAP while requiring signifi One of the most significant advantages of choosing YOLO26 is the well-maintained ecosystem provided by Ultralytics. While YOLOX requires navigating complex research codebases and manual environment setups, Ultralytics offers a streamlined, "zero-to-hero" developer experience. -Using the unified Python API, developers can easily switch between tasks such as [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/). YOLOX, conversely, is strictly limited to bounding box detection. +Using the unified Python API, developers can easily switch between tasks such as [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose). YOLOX, conversely, is strictly limited to bounding box detection. ### Training Example @@ -108,15 +108,15 @@ Choosing the right model dictates the success of your real-world deployment. ### Edge AI and IoT -For applications requiring local processing on limited hardware, such as smart [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/) or remote environmental sensors, **YOLO26** is the definitive choice. Its NMS-free architecture and 43% faster CPU execution mean it runs smoothly on devices like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) without complex quantization workarounds. +For applications requiring local processing on limited hardware, such as smart [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system) or remote environmental sensors, **YOLO26** is the definitive choice. Its NMS-free architecture and 43% faster CPU execution mean it runs smoothly on devices like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) without complex quantization workarounds. ### Autonomous Robotics -Robotics require high precision and low latency. The [pose estimation](https://docs.ultralytics.com/tasks/pose/) capabilities of YOLO26, bolstered by Residual Log-Likelihood Estimation (RLE), allow robots to understand human kinematics in real-time. YOLOX's lack of native keypoint detection makes it unsuitable for such advanced human-robot interaction tasks. +Robotics require high precision and low latency. The [pose estimation](https://docs.ultralytics.com/tasks/pose) capabilities of YOLO26, bolstered by Residual Log-Likelihood Estimation (RLE), allow robots to understand human kinematics in real-time. YOLOX's lack of native keypoint detection makes it unsuitable for such advanced human-robot interaction tasks. ### High-Altitude and Aerial Inspection -When inspecting infrastructure via drones, detecting minute defects is paramount. The ProgLoss and STAL functions in YOLO26 drastically improve recall on tiny objects. Additionally, YOLO26 natively supports [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), complete with a specialized angle loss to resolve boundary issues, making it perfect for satellite and aerial imagery where objects are arbitrarily rotated. +When inspecting infrastructure via drones, detecting minute defects is paramount. The ProgLoss and STAL functions in YOLO26 drastically improve recall on tiny objects. Additionally, YOLO26 natively supports [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb), complete with a specialized angle loss to resolve boundary issues, making it perfect for satellite and aerial imagery where objects are arbitrarily rotated. ### Legacy Deployments @@ -124,6 +124,6 @@ When inspecting infrastructure via drones, detecting minute defects is paramount ## Exploring Other Models -While YOLO26 represents the current state-of-the-art, the Ultralytics ecosystem offers a variety of models tailored to specific needs. For developers interested in transformer-based architectures, [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) provides an alternative approach to end-to-end detection. Additionally, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a robust, highly tested option for production environments that require extensive historical benchmarking. +While YOLO26 represents the current state-of-the-art, the Ultralytics ecosystem offers a variety of models tailored to specific needs. For developers interested in transformer-based architectures, [RT-DETR](https://docs.ultralytics.com/models/rtdetr) provides an alternative approach to end-to-end detection. Additionally, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a robust, highly tested option for production environments that require extensive historical benchmarking. In summary, the transition from YOLOX to YOLO26 illustrates the rapid advancement of the field. By combining an intuitive API, a versatile feature set, and unparalleled efficiency, YOLO26 stands as the premier choice for researchers and developers worldwide. diff --git a/docs/en/compare/yolov10-vs-damo-yolo.md b/docs/en/compare/yolov10-vs-damo-yolo.md index a4814b0559a..fc29fa3fff6 100644 --- a/docs/en/compare/yolov10-vs-damo-yolo.md +++ b/docs/en/compare/yolov10-vs-damo-yolo.md @@ -17,7 +17,7 @@ Whether your project requires deployment on constrained [edge AI](https://www.ul ## Exploring YOLOv10 -Introduced by researchers at Tsinghua University, [YOLOv10](https://docs.ultralytics.com/models/yolov10/) revolutionized the YOLO family by introducing a natively end-to-end approach, effectively eliminating the need for Non-Maximum Suppression (NMS) during post-processing. +Introduced by researchers at Tsinghua University, [YOLOv10](https://docs.ultralytics.com/models/yolov10) revolutionized the YOLO family by introducing a natively end-to-end approach, effectively eliminating the need for Non-Maximum Suppression (NMS) during post-processing. **YOLOv10 Details:** @@ -26,7 +26,7 @@ Introduced by researchers at Tsinghua University, [YOLOv10](https://docs.ultraly - Date: 2024-05-23 - Arxiv: [https://arxiv.org/abs/2405.14458](https://arxiv.org/abs/2405.14458) - GitHub: [https://github.com/THU-MIG/yolov10](https://github.com/THU-MIG/yolov10) -- Docs: [https://docs.ultralytics.com/models/yolov10/](https://docs.ultralytics.com/models/yolov10/) +- Docs: [https://docs.ultralytics.com/models/yolov10/](https://docs.ultralytics.com/models/yolov10) ### Key Architectural Features @@ -36,13 +36,13 @@ Furthermore, the model employs a **Holistic Efficiency-Accuracy Driven Design**. !!! tip "Streamlined Export for Production" - Because YOLOv10 removes NMS operations from the inference graph, exporting the model to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/) or [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) is highly simplified, making it exceptionally well-suited for edge deployments. + Because YOLOv10 removes NMS operations from the inference graph, exporting the model to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx) or [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) is highly simplified, making it exceptionally well-suited for edge deployments. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ### Usage Example -YOLOv10 is deeply integrated into the Ultralytics ecosystem, making it incredibly easy to use via the [Ultralytics Python package](https://docs.ultralytics.com/quickstart/). +YOLOv10 is deeply integrated into the Ultralytics ecosystem, making it incredibly easy to use via the [Ultralytics Python package](https://docs.ultralytics.com/quickstart). ```python from ultralytics import YOLO @@ -122,11 +122,11 @@ DAMO-YOLO is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage @@ -138,15 +138,15 @@ Unlike standalone academic repositories that often face abandonment, Ultralytics ### Training Efficiency and Memory Requirements -Training large models can quickly become computationally expensive. The Ultralytics YOLO architectures are historically known for their low CUDA memory footprint during training and inference. This efficiency allows developers to train models on consumer-grade hardware or cost-effective cloud instances without running into out-of-memory errors that are common when working with transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +Training large models can quickly become computationally expensive. The Ultralytics YOLO architectures are historically known for their low CUDA memory footprint during training and inference. This efficiency allows developers to train models on consumer-grade hardware or cost-effective cloud instances without running into out-of-memory errors that are common when working with transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). !!! note "Experiment Tracking" - Ultralytics natively integrates with top MLOps tools. You can easily track your model training progress using integrations with [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/), [Comet](https://docs.ultralytics.com/integrations/comet/), or [ClearML](https://docs.ultralytics.com/integrations/clearml/) with zero additional boilerplate code. + Ultralytics natively integrates with top MLOps tools. You can easily track your model training progress using integrations with [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases), [Comet](https://docs.ultralytics.com/integrations/comet), or [ClearML](https://docs.ultralytics.com/integrations/clearml) with zero additional boilerplate code. ### Versatility Across Tasks -A significant limitation of many specialized detection models is their narrow focus. Within the Ultralytics ecosystem, you are not limited to just object detection. The tools seamlessly extend to multiple [computer vision tasks](https://docs.ultralytics.com/tasks/), including [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding box (OBB) detection](https://docs.ultralytics.com/tasks/obb/). +A significant limitation of many specialized detection models is their narrow focus. Within the Ultralytics ecosystem, you are not limited to just object detection. The tools seamlessly extend to multiple [computer vision tasks](https://docs.ultralytics.com/tasks), including [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding box (OBB) detection](https://docs.ultralytics.com/tasks/obb). ## Looking Ahead: The YOLO26 Evolution @@ -161,4 +161,4 @@ Key advancements in YOLO26 include: - **MuSGD Optimizer:** A hybrid of SGD and Muon, bringing advanced LLM training stability and faster convergence directly into computer vision. - **ProgLoss + STAL:** Drastically improved loss functions that offer notable enhancements in small-object recognition, which is essential for use cases like [agriculture](https://www.ultralytics.com/solutions/ai-in-agriculture) and remote sensing. -By utilizing the newly revamped [Ultralytics Platform](https://docs.ultralytics.com/platform/), developers can seamlessly annotate, train, and deploy next-generation models like YOLO26 in just a few clicks, ensuring your computer vision pipeline is both cutting-edge and future-proof. +By utilizing the newly revamped [Ultralytics Platform](https://docs.ultralytics.com/platform), developers can seamlessly annotate, train, and deploy next-generation models like YOLO26 in just a few clicks, ensuring your computer vision pipeline is both cutting-edge and future-proof. diff --git a/docs/en/compare/yolov10-vs-efficientdet.md b/docs/en/compare/yolov10-vs-efficientdet.md index fd36f7d5d95..7b270f8be99 100644 --- a/docs/en/compare/yolov10-vs-efficientdet.md +++ b/docs/en/compare/yolov10-vs-efficientdet.md @@ -6,7 +6,7 @@ keywords: YOLOv10, EfficientDet, object detection, model comparison, real-time d # YOLOv10 vs EfficientDet: Comparing Real-Time Object Detection Architectures -Selecting the optimal neural network for [object detection](https://docs.ultralytics.com/tasks/detect/) is a critical decision that dictates the success of modern computer vision systems. Two prominent architectures that have significantly influenced the field are **YOLOv10** and **EfficientDet**. While both aim to maximize accuracy while minimizing computational overhead, they take vastly different architectural approaches to achieve these goals. +Selecting the optimal neural network for [object detection](https://docs.ultralytics.com/tasks/detect) is a critical decision that dictates the success of modern computer vision systems. Two prominent architectures that have significantly influenced the field are **YOLOv10** and **EfficientDet**. While both aim to maximize accuracy while minimizing computational overhead, they take vastly different architectural approaches to achieve these goals. This comprehensive guide dives into their unique designs, training methodologies, and deployment characteristics, helping developers and ML engineers make data-driven decisions for [vision AI applications](https://www.ultralytics.com/blog/exploring-various-types-of-data-for-vision-ai-applications). We will examine how they perform on hardware ranging from embedded [edge AI devices](https://www.ultralytics.com/glossary/edge-ai) to powerful cloud GPUs. @@ -30,7 +30,7 @@ YOLOv10 introduces consistent dual assignments for NMS-free training. During tra - **Date:** 2024-05-23 - **Paper:** [YOLOv10: Real-Time End-to-End Object Detection](https://arxiv.org/abs/2405.14458) - **GitHub:** [THU-MIG/yolov10](https://github.com/THU-MIG/yolov10) -- **Docs:** [YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10/) +- **Docs:** [YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10) !!! tip "Streamlined Deployment" @@ -39,14 +39,14 @@ YOLOv10 introduces consistent dual assignments for NMS-free training. During tra **Strengths:** - **Predictable Inference:** The removal of NMS ensures consistent inference times regardless of the number of objects in the scene. -- **Lower Memory Usage:** Compared to transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/), YOLOv10 enjoys significantly lower memory requirements during both training and inference. -- **Excellent Speed/Accuracy Trade-off:** Specifically optimized for low-latency scenarios without sacrificing [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/). +- **Lower Memory Usage:** Compared to transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr), YOLOv10 enjoys significantly lower memory requirements during both training and inference. +- **Excellent Speed/Accuracy Trade-off:** Specifically optimized for low-latency scenarios without sacrificing [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics). **Weaknesses:** -- **Single Task Focus:** Unlike the broader [Ultralytics ecosystem](https://docs.ultralytics.com/), the original YOLOv10 repository is heavily focused on detection, lacking native support for [instance segmentation](https://docs.ultralytics.com/tasks/segment/) or [pose estimation](https://docs.ultralytics.com/tasks/pose/). +- **Single Task Focus:** Unlike the broader [Ultralytics ecosystem](https://docs.ultralytics.com/), the original YOLOv10 repository is heavily focused on detection, lacking native support for [instance segmentation](https://docs.ultralytics.com/tasks/segment) or [pose estimation](https://docs.ultralytics.com/tasks/pose). -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## EfficientDet: Scalable and Balanced @@ -123,11 +123,11 @@ EfficientDet is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Modern Standard: Enter Ultralytics YOLO26 @@ -137,11 +137,11 @@ While YOLOv10 introduced the groundbreaking NMS-free paradigm and EfficientDet s 1. **End-to-End NMS-Free Design:** YOLO26 natively adopts the end-to-end NMS-free architecture pioneered in YOLOv10, streamlining deployment and accelerating inference. 2. **Up to 43% Faster CPU Inference:** For edge devices lacking dedicated accelerators, YOLO26 is specifically optimized to run efficiently on standard CPUs. -3. **Advanced MuSGD Optimizer:** Inspired by LLM training innovations, YOLO26 utilizes a hybrid of SGD and Muon for incredibly stable training and rapid convergence, vastly improving [training efficiency](https://docs.ultralytics.com/guides/model-training-tips/) compared to EfficientDet. +3. **Advanced MuSGD Optimizer:** Inspired by LLM training innovations, YOLO26 utilizes a hybrid of SGD and Muon for incredibly stable training and rapid convergence, vastly improving [training efficiency](https://docs.ultralytics.com/guides/model-training-tips) compared to EfficientDet. 4. **ProgLoss + STAL:** These improved loss functions deliver remarkable boosts in small-object recognition, a traditional weak point for both YOLOv10 and EfficientDet. 5. **DFL Removal:** By removing Distribution Focal Loss, YOLO26 exports seamlessly to nearly any hardware format, including [OpenVINO](https://docs.openvino.ai/) and CoreML. -Furthermore, YOLO26 provides unmatched **versatility**. While EfficientDet and YOLOv10 are strictly detection models, YOLO26 seamlessly handles [oriented bounding boxes](https://docs.ultralytics.com/tasks/obb/), [image classification](https://docs.ultralytics.com/tasks/classify/), and instance segmentation using the same intuitive [Ultralytics Python package](https://docs.ultralytics.com/usage/python/). +Furthermore, YOLO26 provides unmatched **versatility**. While EfficientDet and YOLOv10 are strictly detection models, YOLO26 seamlessly handles [oriented bounding boxes](https://docs.ultralytics.com/tasks/obb), [image classification](https://docs.ultralytics.com/tasks/classify), and instance segmentation using the same intuitive [Ultralytics Python package](https://docs.ultralytics.com/usage/python). !!! tip "Well-Maintained Ecosystem" @@ -149,7 +149,7 @@ Furthermore, YOLO26 provides unmatched **versatility**. While EfficientDet and Y ### Ease of Use with Ultralytics -The well-maintained ecosystem provided by Ultralytics ensures a smooth developer experience. Training a model, validating it, and exporting it to [TensorRT integration](https://docs.ultralytics.com/integrations/tensorrt/) takes only a few lines of code. +The well-maintained ecosystem provided by Ultralytics ensures a smooth developer experience. Training a model, validating it, and exporting it to [TensorRT integration](https://docs.ultralytics.com/integrations/tensorrt) takes only a few lines of code. ```python from ultralytics import YOLO diff --git a/docs/en/compare/yolov10-vs-pp-yoloe.md b/docs/en/compare/yolov10-vs-pp-yoloe.md index 2ddfc0ab61c..97fca182978 100644 --- a/docs/en/compare/yolov10-vs-pp-yoloe.md +++ b/docs/en/compare/yolov10-vs-pp-yoloe.md @@ -8,7 +8,7 @@ keywords: YOLOv10,PP-YOLOE+,object detection,model comparison,computer vision,Ul In the rapidly evolving landscape of [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv), choosing the optimal architecture for real-time object detection is crucial for balancing accuracy, inference speed, and deployment efficiency. Two notable contenders in this arena are **YOLOv10** and **PP-YOLOE+**. While both models offer robust capabilities, they originate from different design philosophies and ecosystem integrations. -This technical guide provides an in-depth analysis of these two architectures, exploring their [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/), structural differences, and ideal real-world applications. By understanding the nuances of each, machine learning engineers and researchers can make informed decisions for their deployment pipelines. +This technical guide provides an in-depth analysis of these two architectures, exploring their [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics), structural differences, and ideal real-world applications. By understanding the nuances of each, machine learning engineers and researchers can make informed decisions for their deployment pipelines. @@ -26,15 +26,15 @@ Developed by researchers at Tsinghua University, YOLOv10 introduced a significan - **Date:** 2024-05-23 - **Arxiv:** [2405.14458](https://arxiv.org/abs/2405.14458) - **GitHub:** [THU-MIG/yolov10](https://github.com/THU-MIG/yolov10) -- **Docs:** [YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10/) +- **Docs:** [YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10) ### Architectural Strengths and Weaknesses YOLOv10's standout feature is its consistent dual assignments for NMS-free training, which allows it to predict bounding boxes directly without relying on heuristic thresholding. This results in an excellent balance of speed and precision, particularly for the smaller model variants. The architecture also employs a holistic efficiency-accuracy driven design, minimizing computational redundancy. -However, as a strictly detection-focused model, it lacks the native versatility found in models that support [instance segmentation](https://docs.ultralytics.com/tasks/segment/) or [pose estimation](https://docs.ultralytics.com/tasks/pose/) out of the box. +However, as a strictly detection-focused model, it lacks the native versatility found in models that support [instance segmentation](https://docs.ultralytics.com/tasks/segment) or [pose estimation](https://docs.ultralytics.com/tasks/pose) out of the box. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## PP-YOLOE+: The PaddlePaddle Powerhouse @@ -53,7 +53,7 @@ PP-YOLOE+ is an upgraded version of the original PP-YOLOE, developed by Baidu's PP-YOLOE+ utilizes a scalable backbone and a powerful neck design (CSPRepResNet) that significantly boosts feature extraction. Its training methodology relies heavily on large-scale datasets like Objects365 for pre-training, which contributes to its impressive accuracy, particularly on the larger `x` and `l` variants. -The primary drawback of PP-YOLOE+ is its deep entanglement with the PaddlePaddle framework. For teams accustomed to PyTorch or the unified Ultralytics ecosystem, adopting PP-YOLOE+ can introduce friction. Furthermore, its larger parameter count leads to higher memory requirements during training compared to equivalent [Ultralytics YOLO models](https://docs.ultralytics.com/models/). +The primary drawback of PP-YOLOE+ is its deep entanglement with the PaddlePaddle framework. For teams accustomed to PyTorch or the unified Ultralytics ecosystem, adopting PP-YOLOE+ can introduce friction. Furthermore, its larger parameter count leads to higher memory requirements during training compared to equivalent [Ultralytics YOLO models](https://docs.ultralytics.com/models). [Learn more about PP-YOLOE+](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.8.1/configs/ppyoloe){ .md-button } @@ -100,11 +100,11 @@ PP-YOLOE+ is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage and the Future: YOLO26 @@ -117,13 +117,13 @@ While YOLOv10 and PP-YOLOE+ offer specialized benefits, the modern standard for ### Key Innovations in YOLO26 - **End-to-End NMS-Free Design:** By eliminating post-processing latency, YOLO26 guarantees stable, high-speed inferences, vital for [autonomous vehicles](https://www.ultralytics.com/glossary/autonomous-vehicles) and rapid robotics. -- **Edge-First Optimizations:** The removal of Distribution Focal Loss (DFL) simplifies model [export formats](https://docs.ultralytics.com/modes/export/) and yields up to **43% faster CPU inference** over previous generations. +- **Edge-First Optimizations:** The removal of Distribution Focal Loss (DFL) simplifies model [export formats](https://docs.ultralytics.com/modes/export) and yields up to **43% faster CPU inference** over previous generations. - **Advanced Training Dynamics:** Leveraging the new **MuSGD Optimizer**—a hybrid of SGD and Muon—YOLO26 brings LLM training stability to vision tasks, converging faster and more reliably. - **Enhanced Accuracy via ProgLoss + STAL:** These advanced loss functions specifically target complex scenarios, offering exceptional gains in small-object detection crucial for [aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision) and [agriculture](https://www.ultralytics.com/solutions/ai-in-agriculture). ### Unmatched Versatility -Unlike PP-YOLOE+ which focuses on detection, YOLO26 handles [image classification](https://docs.ultralytics.com/tasks/classify/), [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), pose estimation, and segmentation from a single, unified codebase. You can easily manage [datasets](https://docs.ultralytics.com/datasets/), train, and deploy models directly via the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolo26). +Unlike PP-YOLOE+ which focuses on detection, YOLO26 handles [image classification](https://docs.ultralytics.com/tasks/classify), [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb), pose estimation, and segmentation from a single, unified codebase. You can easily manage [datasets](https://docs.ultralytics.com/datasets), train, and deploy models directly via the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolo26). ```python from ultralytics import YOLO @@ -144,8 +144,8 @@ Selecting the right model heavily depends on deployment constraints: - **PP-YOLOE+** shines in specific industrial deployments across Asia where the Baidu hardware-software stack is pre-established. It handles static, high-resolution [quality inspection in manufacturing](https://www.ultralytics.com/blog/quality-inspection-in-manufacturing-traditional-vs-deep-learning-methods) well. - **YOLOv10** is optimal for dense [crowd management](https://www.ultralytics.com/blog/vision-ai-in-crowd-management) and environments where removing NMS drops latency variability, making real-time tracking more consistent. -- **Ultralytics YOLO26** remains the definitive choice for enterprise-wide scaling. Whether analyzing traffic in [smart cities](https://www.ultralytics.com/blog/computer-vision-ai-in-smart-cities) or deploying to ultra-low-power edge nodes like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/), its minimal memory footprint, comprehensive documentation, and unified training pipeline ensure rapid ROI. +- **Ultralytics YOLO26** remains the definitive choice for enterprise-wide scaling. Whether analyzing traffic in [smart cities](https://www.ultralytics.com/blog/computer-vision-ai-in-smart-cities) or deploying to ultra-low-power edge nodes like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi), its minimal memory footprint, comprehensive documentation, and unified training pipeline ensure rapid ROI. -For those interested in exploring older supported architectures or transformer alternatives within the ecosystem, see the documentations for [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +For those interested in exploring older supported architectures or transformer alternatives within the ecosystem, see the documentations for [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or [RT-DETR](https://docs.ultralytics.com/models/rtdetr). Ultimately, a well-maintained ecosystem combined with a simple API ensures that developers spend less time debugging configuration files and more time solving real-world [vision AI](https://www.ultralytics.com/blog-category/vision-ai) problems. diff --git a/docs/en/compare/yolov10-vs-rtdetr.md b/docs/en/compare/yolov10-vs-rtdetr.md index a0872dc19d2..1a2d4999bbe 100644 --- a/docs/en/compare/yolov10-vs-rtdetr.md +++ b/docs/en/compare/yolov10-vs-rtdetr.md @@ -26,17 +26,17 @@ Developed by researchers at Tsinghua University, YOLOv10 focuses heavily on arch - Date: 2024-05-23 - ArXiv: [YOLOv10 Paper](https://arxiv.org/abs/2405.14458) - GitHub: [THU-MIG/yolov10](https://github.com/THU-MIG/yolov10) -- Docs: [YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10/) +- Docs: [YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10) ### Architecture and Methodologies -YOLOv10's primary breakthrough is its holistic efficiency-accuracy driven model design. It optimizes various components from both perspectives, greatly reducing computational overhead. The consistent dual assignments strategy allows the model to train without relying on NMS, which translates to a streamlined, end-to-end deployment pipeline. This is particularly beneficial when exporting models to edge formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/) or [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), where post-processing operations can introduce unexpected latency. +YOLOv10's primary breakthrough is its holistic efficiency-accuracy driven model design. It optimizes various components from both perspectives, greatly reducing computational overhead. The consistent dual assignments strategy allows the model to train without relying on NMS, which translates to a streamlined, end-to-end deployment pipeline. This is particularly beneficial when exporting models to edge formats like [ONNX](https://docs.ultralytics.com/integrations/onnx) or [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), where post-processing operations can introduce unexpected latency. ### Strengths and Weaknesses -The model boasts exceptional speed-accuracy trade-offs, especially in the smaller variants (N and S). Its minimal latency makes it ideal for high-speed edge environments. However, while YOLOv10 excels at raw detection speed, it remains a specialized detection-only model. Teams requiring [instance segmentation](https://docs.ultralytics.com/tasks/segment/) or [pose estimation](https://docs.ultralytics.com/tasks/pose/) will need to look towards more versatile frameworks. +The model boasts exceptional speed-accuracy trade-offs, especially in the smaller variants (N and S). Its minimal latency makes it ideal for high-speed edge environments. However, while YOLOv10 excels at raw detection speed, it remains a specialized detection-only model. Teams requiring [instance segmentation](https://docs.ultralytics.com/tasks/segment) or [pose estimation](https://docs.ultralytics.com/tasks/pose) will need to look towards more versatile frameworks. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## RTDETRv2: Refining the Detection Transformer @@ -49,7 +49,7 @@ Building upon the original Real-Time Detection Transformer, RTDETRv2 incorporate - Date: 2024-07-24 - ArXiv: [RTDETRv2 Paper](https://arxiv.org/abs/2407.17140) - GitHub: [lyuwenyu/RT-DETR](https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetrv2_pytorch) -- Docs: [RTDETRv2 Documentation](https://docs.ultralytics.com/models/rtdetr/) +- Docs: [RTDETRv2 Documentation](https://docs.ultralytics.com/models/rtdetr) ### Architecture and Methodologies @@ -59,11 +59,11 @@ RTDETRv2 utilizes a hybrid architecture, combining a Convolutional Neural Networ The transformer architecture provides excellent accuracy, particularly on larger parameter scales, and natively outputs final detections without NMS. However, this comes at a cost. Transformer models traditionally require significantly more CUDA memory during training and can be slower to converge compared to pure CNN architectures. While RTDETRv2 has improved inference speeds, it generally consumes more memory than lightweight YOLO variants. -[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETRv2](https://docs.ultralytics.com/models/rtdetr){ .md-button } ## Performance Comparison -Evaluating the performance metrics provides a clearer picture of where each model excels. The following table highlights their capabilities on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/): +Evaluating the performance metrics provides a clearer picture of where each model excels. The following table highlights their capabilities on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco): | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -103,11 +103,11 @@ RT-DETR is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Ecosystem and Innovation @@ -121,7 +121,7 @@ For developers seeking the absolute best performance, [Ultralytics YOLO26](https YOLO26 brings LLM training innovations to computer vision via the **MuSGD Optimizer** (a hybrid of SGD and Muon), resulting in more stable training and faster convergence. It also boasts up to **43% Faster CPU Inference**, making it the premier choice for edge computing. -Furthermore, YOLO26 introduces **ProgLoss + STAL** for notable improvements in small-object recognition, and unlike the specialized YOLOv10, it offers extreme versatility. It natively supports [object detection](https://docs.ultralytics.com/tasks/detect/), segmentation, pose, and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) with task-specific improvements like semantic segmentation loss and Residual Log-Likelihood Estimation (RLE) for pose. Furthermore, the removal of Distribution Focal Loss (DFL) ensures simplified export and better low-power device compatibility. +Furthermore, YOLO26 introduces **ProgLoss + STAL** for notable improvements in small-object recognition, and unlike the specialized YOLOv10, it offers extreme versatility. It natively supports [object detection](https://docs.ultralytics.com/tasks/detect), segmentation, pose, and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb) with task-specific improvements like semantic segmentation loss and Residual Log-Likelihood Estimation (RLE) for pose. Furthermore, the removal of Distribution Focal Loss (DFL) ensures simplified export and better low-power device compatibility. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -141,7 +141,7 @@ model_rtdetr = RTDETR("rtdetr-l.pt") results = model_rtdetr.predict("https://ultralytics.com/images/bus.jpg") ``` -The well-maintained ecosystem provides tools for easy [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/) and integrates flawlessly with extensive tracking solutions and [model deployment options](https://docs.ultralytics.com/guides/model-deployment-options/). +The well-maintained ecosystem provides tools for easy [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning) and integrates flawlessly with extensive tracking solutions and [model deployment options](https://docs.ultralytics.com/guides/model-deployment-options). ## Conclusion diff --git a/docs/en/compare/yolov10-vs-yolo11.md b/docs/en/compare/yolov10-vs-yolo11.md index 196e2732966..4561b14dff3 100644 --- a/docs/en/compare/yolov10-vs-yolo11.md +++ b/docs/en/compare/yolov10-vs-yolo11.md @@ -19,11 +19,11 @@ Released in the spring of 2024, YOLOv10 introduced a novel approach to the tradi - **Date:** May 23, 2024 - **Research Paper:** [arXiv:2405.14458](https://arxiv.org/abs/2405.14458) - **Source Code:** [THU-MIG/yolov10 on GitHub](https://github.com/THU-MIG/yolov10) -- **Documentation:** [YOLOv10 Docs](https://docs.ultralytics.com/models/yolov10/) +- **Documentation:** [YOLOv10 Docs](https://docs.ultralytics.com/models/yolov10) The standout innovation of YOLOv10 is its consistent dual assignments strategy, which enables NMS-free training. Traditional object detectors rely heavily on [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) to filter out redundant bounding box predictions. By removing this step, YOLOv10 achieves true end-to-end detection, reducing inference latency and simplifying deployment on hardware accelerators like [Neural Processing Units (NPUs)](https://en.wikipedia.org/wiki/AI_accelerator) where custom NMS operations are notoriously difficult to optimize. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## YOLO11: Ecosystem-Driven Versatility and Performance @@ -35,20 +35,20 @@ Launched later in the same year, YOLO11 represents the continuous refinement of - **Source Code:** [Ultralytics on GitHub](https://github.com/ultralytics/ultralytics) - **Platform Integration:** [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolo11) -YOLO11 is designed for production. While it excels at standard bounding box detection, its true strength lies in its **versatility**. Unlike YOLOv10, which is primarily focused on object detection, YOLO11 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) tasks using a unified architecture. It boasts remarkably low **memory requirements** during training, making it highly accessible for teams working with consumer-grade [GPUs](https://www.ultralytics.com/glossary/gpu-graphics-processing-unit) compared to heavier, transformer-based architectures. +YOLO11 is designed for production. While it excels at standard bounding box detection, its true strength lies in its **versatility**. Unlike YOLOv10, which is primarily focused on object detection, YOLO11 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) tasks using a unified architecture. It boasts remarkably low **memory requirements** during training, making it highly accessible for teams working with consumer-grade [GPUs](https://www.ultralytics.com/glossary/gpu-graphics-processing-unit) compared to heavier, transformer-based architectures. [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } ## Performance and Metrics Comparison -When comparing these models side-by-side, it is essential to look at how they perform across different scale variants on standard benchmarks like the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). +When comparing these models side-by-side, it is essential to look at how they perform across different scale variants on standard benchmarks like the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). -The table below highlights the performance differences. YOLO11 frequently edges out YOLOv10 in mAP across most size categories while maintaining highly competitive [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) inference speeds. +The table below highlights the performance differences. YOLO11 frequently edges out YOLOv10 in mAP across most size categories while maintaining highly competitive [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) inference speeds. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | -------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -67,7 +67,7 @@ The table below highlights the performance differences. YOLO11 frequently edges !!! tip "Hardware Acceleration" - To reproduce these rapid inference speeds locally, ensure you export your models to optimized formats like [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) for Intel CPUs or TensorRT for NVIDIA GPUs. + To reproduce these rapid inference speeds locally, ensure you export your models to optimized formats like [OpenVINO](https://docs.ultralytics.com/integrations/openvino) for Intel CPUs or TensorRT for NVIDIA GPUs. ## Architectural Deep Dive @@ -75,13 +75,13 @@ The table below highlights the performance differences. YOLO11 frequently edges YOLOv10's architecture emphasizes reducing computational redundancy. By optimizing the backbone and neck designs using a holistic efficiency-accuracy driven strategy, the authors from Tsinghua University managed to lower the parameter count significantly in the mid-tier models (like YOLOv10m) compared to previous iterations. -However, **Training Efficiency** is a major hallmark of Ultralytics models. YOLO11 utilizes the highly refined `ultralytics` Python package, which abstracts away complex [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/). This framework automatically handles advanced data augmentations, learning rate scheduling, and multi-GPU distributed training out of the box. YOLO11's architecture also exhibits excellent gradient flow, resulting in faster convergence and lower VRAM usage during the training phase. +However, **Training Efficiency** is a major hallmark of Ultralytics models. YOLO11 utilizes the highly refined `ultralytics` Python package, which abstracts away complex [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning). This framework automatically handles advanced data augmentations, learning rate scheduling, and multi-GPU distributed training out of the box. YOLO11's architecture also exhibits excellent gradient flow, resulting in faster convergence and lower VRAM usage during the training phase. ### Ease of Use and The Ecosystem Advantage A critical factor for enterprise adoption is the **Well-Maintained Ecosystem**. Research repositories, while groundbreaking, often become dormant after the initial paper publication. The Ultralytics ecosystem, backing YOLO11, provides a seamless, end-to-end developer experience. -Integrating seamlessly with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) for experiment tracking and [Roboflow](https://docs.ultralytics.com/integrations/roboflow/) for dataset management, YOLO11 accelerates the transition from prototype to production. The **Ease of Use** is evident in the streamlined API, allowing developers to train and export models with just a few lines of code. +Integrating seamlessly with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) for experiment tracking and [Roboflow](https://docs.ultralytics.com/integrations/roboflow) for dataset management, YOLO11 accelerates the transition from prototype to production. The **Ease of Use** is evident in the streamlined API, allowing developers to train and export models with just a few lines of code. ```python from ultralytics import YOLO @@ -112,21 +112,21 @@ YOLOv10 is a strong choice for: YOLO11 is recommended for: -- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) where reliability and active maintenance are paramount. -- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [OBB](https://docs.ultralytics.com/tasks/obb/) within a single unified framework. -- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python/). +- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) where reliability and active maintenance are paramount. +- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [OBB](https://docs.ultralytics.com/tasks/obb) within a single unified framework. +- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python). ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Exploring Other Architectures -While YOLOv10 and YOLO11 are excellent choices, your specific use case might benefit from other architectures available in the documentation. For sequence-based reasoning, transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) provide high accuracy, though they typically demand higher memory requirements. Conversely, if you need zero-shot capabilities for identifying novel classes without retraining, [YOLO-World](https://docs.ultralytics.com/models/yolo-world/) offers an open-vocabulary approach driven by natural language prompts. +While YOLOv10 and YOLO11 are excellent choices, your specific use case might benefit from other architectures available in the documentation. For sequence-based reasoning, transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) provide high accuracy, though they typically demand higher memory requirements. Conversely, if you need zero-shot capabilities for identifying novel classes without retraining, [YOLO-World](https://docs.ultralytics.com/models/yolo-world) offers an open-vocabulary approach driven by natural language prompts. ## The Next Generation: YOLO26 diff --git a/docs/en/compare/yolov10-vs-yolo26.md b/docs/en/compare/yolov10-vs-yolo26.md index ba8afecd85e..e3b2183b25e 100644 --- a/docs/en/compare/yolov10-vs-yolo26.md +++ b/docs/en/compare/yolov10-vs-yolo26.md @@ -24,7 +24,7 @@ Released in mid-2024, YOLOv10 represented a significant leap forward in academic The Tsinghua University team introduced a consistent dual assignment strategy for NMS-free training. This allowed the model to predict bounding boxes accurately without requiring a post-processing filtering step, directly improving inference latency and lowering the barrier for deployment on hardware accelerators. While highly efficient for standard detection tasks, the model primarily focused on bounding box prediction and lacked native support for more complex tasks like instance segmentation or pose estimation. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## YOLO26: The New Standard for Edge and Cloud Vision AI @@ -47,7 +47,7 @@ Furthermore, training stability and convergence speed have been revolutionized t ## Technical Performance Comparison -When evaluating these models, the balance between accuracy, model size, and inference speed is critical. The table below highlights the performance of both model families across various scales, evaluated on the standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). +When evaluating these models, the balance between accuracy, model size, and inference speed is critical. The table below highlights the performance of both model families across various scales, evaluated on the standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | -------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -70,7 +70,7 @@ The data clearly demonstrates the evolutionary advantage of the newer architectu A model is only as useful as the ecosystem supporting it. While YOLOv10 provided an excellent academic implementation based on [PyTorch](https://pytorch.org/), it often requires manual configuration for tasks beyond basic detection. -In contrast, YOLO26 is fully integrated into the well-maintained Ultralytics ecosystem. This ensures significantly lower memory requirements during training compared to transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/), allowing researchers to train state-of-the-art networks on consumer-grade hardware. The ease of use is unparalleled, offering a unified API that handles data augmentation, hyperparameter tuning, and logging automatically. +In contrast, YOLO26 is fully integrated into the well-maintained Ultralytics ecosystem. This ensures significantly lower memory requirements during training compared to transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr), allowing researchers to train state-of-the-art networks on consumer-grade hardware. The ease of use is unparalleled, offering a unified API that handles data augmentation, hyperparameter tuning, and logging automatically. ### Code Example: Training YOLO26 @@ -99,15 +99,15 @@ Choosing the right architecture depends entirely on deployment constraints. ### High-Speed Edge Computing -For applications requiring rapid deployment on microcontrollers, robotics, or legacy mobile devices, the 43% faster CPU inference of YOLO26 makes it the definitive choice. Its NMS-free, DFL-free architecture converts seamlessly to formats like [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), ideal for real-time video analytics in smart city infrastructure. +For applications requiring rapid deployment on microcontrollers, robotics, or legacy mobile devices, the 43% faster CPU inference of YOLO26 makes it the definitive choice. Its NMS-free, DFL-free architecture converts seamlessly to formats like [OpenVINO](https://docs.ultralytics.com/integrations/openvino) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), ideal for real-time video analytics in smart city infrastructure. ### Advanced Multi-Task Vision -While YOLOv10 excels in pure bounding box detection, projects requiring rich visual understanding must rely on YOLO26. From [instance segmentation](https://docs.ultralytics.com/tasks/segment/) in medical imaging to precision [pose estimation](https://docs.ultralytics.com/tasks/pose/) for sports analytics, YOLO26 provides task-specific loss functions that guarantee superior accuracy across diverse domains. +While YOLOv10 excels in pure bounding box detection, projects requiring rich visual understanding must rely on YOLO26. From [instance segmentation](https://docs.ultralytics.com/tasks/segment) in medical imaging to precision [pose estimation](https://docs.ultralytics.com/tasks/pose) for sports analytics, YOLO26 provides task-specific loss functions that guarantee superior accuracy across diverse domains. !!! note "Alternative Options" - If your project requires robust open-vocabulary detection, consider exploring [YOLO-World](https://docs.ultralytics.com/models/yolo-world/). For users maintaining legacy pipelines, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a fully supported and powerful alternative within the Ultralytics framework. + If your project requires robust open-vocabulary detection, consider exploring [YOLO-World](https://docs.ultralytics.com/models/yolo-world). For users maintaining legacy pipelines, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a fully supported and powerful alternative within the Ultralytics framework. ## Use Cases and Recommendations @@ -127,7 +127,7 @@ YOLO26 is recommended for: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Conclusion diff --git a/docs/en/compare/yolov10-vs-yolov5.md b/docs/en/compare/yolov10-vs-yolov5.md index 040416ad9a1..78b91fe81e6 100644 --- a/docs/en/compare/yolov10-vs-yolov5.md +++ b/docs/en/compare/yolov10-vs-yolov5.md @@ -29,7 +29,7 @@ Developed by researchers at Tsinghua University, YOLOv10 introduced a novel appr The defining breakthrough of YOLOv10 is its **End-to-End NMS-Free Design**. Historically, YOLO models relied on [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) to filter out redundant bounding boxes. YOLOv10 utilizes consistent dual assignments for NMS-free training, which drastically reduces inference latency variability and simplifies deployment logic. Additionally, the architecture features a holistic efficiency-accuracy driven design that thoroughly optimizes various components to reduce computational redundancy. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ### YOLOv5: The Industry Standard for Usability @@ -40,13 +40,13 @@ Released shortly after the inception of the Ultralytics PyTorch repository, YOLO - **Date:** 2020-06-26 - **Source Code:** [YOLOv5 GitHub Repository](https://github.com/ultralytics/yolov5) -YOLOv5 is celebrated for its **Ease of Use** and highly **Well-Maintained Ecosystem**. Written entirely in PyTorch, it offered a seamless "zero-to-hero" experience with out-of-the-box support for training, validation, and export to formats like [ONNX](https://onnx.ai/) and [TensorRT](https://developer.nvidia.com/tensorrt). Unlike YOLOv10, which focuses primarily on pure object detection, YOLOv5 demonstrates exceptional **Versatility**, supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [image classification](https://docs.ultralytics.com/tasks/classify/) within the same unified Python API. +YOLOv5 is celebrated for its **Ease of Use** and highly **Well-Maintained Ecosystem**. Written entirely in PyTorch, it offered a seamless "zero-to-hero" experience with out-of-the-box support for training, validation, and export to formats like [ONNX](https://onnx.ai/) and [TensorRT](https://developer.nvidia.com/tensorrt). Unlike YOLOv10, which focuses primarily on pure object detection, YOLOv5 demonstrates exceptional **Versatility**, supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [image classification](https://docs.ultralytics.com/tasks/classify) within the same unified Python API. [Learn more about YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5){ .md-button } ## Performance and Metrics Comparison -Visualizing the relationship between speed and accuracy is essential for identifying the models that offer the best accuracy for a given speed constraint. Understanding these [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/) is fundamental to selecting a model that aligns with your specific hardware constraints. +Visualizing the relationship between speed and accuracy is essential for identifying the models that offer the best accuracy for a given speed constraint. Understanding these [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics) is fundamental to selecting a model that aligns with your specific hardware constraints. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | -------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -91,21 +91,21 @@ YOLOv5 is recommended for: - **Proven Production Systems:** Existing deployments where YOLOv5's long track record of stability, extensive documentation, and massive community support are valued. - **Resource-Constrained Training:** Environments with limited GPU resources where YOLOv5's efficient training pipeline and lower memory requirements are advantageous. -- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), [CoreML](https://docs.ultralytics.com/integrations/coreml/), and [TFLite](https://docs.ultralytics.com/integrations/tflite/). +- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), [CoreML](https://docs.ultralytics.com/integrations/coreml), and [TFLite](https://docs.ultralytics.com/integrations/tflite). ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage While YOLOv10 offers excellent detection capabilities, relying on academic repositories can sometimes complicate production pipelines. By using the official [Ultralytics Python package](https://pypi.org/project/ultralytics/), you gain access to a unified ecosystem that supports both YOLOv5 and YOLOv10, along with advanced features. -- **Training Efficiency:** Ultralytics YOLO architectures are deeply optimized for lower [memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics/) during training. Unlike heavy transformer models (such as RT-DETR) which require massive CUDA memory, you can comfortably train YOLOv5 and YOLOv10 on standard consumer GPUs. +- **Training Efficiency:** Ultralytics YOLO architectures are deeply optimized for lower [memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics) during training. Unlike heavy transformer models (such as RT-DETR) which require massive CUDA memory, you can comfortably train YOLOv5 and YOLOv10 on standard consumer GPUs. - **Ecosystem Integration:** The integration with [Ultralytics Platform](https://platform.ultralytics.com) allows developers to visually manage datasets, track experiments using [Weights & Biases](https://wandb.ai/site), and automatically tune hyperparameters. ### Code Example: Seamless Training @@ -140,13 +140,13 @@ If you are starting a new [machine learning](https://www.ultralytics.com/glossar YOLO26 natively incorporates the **End-to-End NMS-Free Design** pioneered by YOLOv10, ensuring rapid, deterministic deployment. Furthermore, YOLO26 introduces several critical breakthroughs: -- **Up to 43% Faster CPU Inference:** By removing the Distribution Focal Loss (DFL) module, YOLO26 achieves massive speedups on standard CPUs, making it the premier choice for [mobile deployment](https://docs.ultralytics.com/guides/model-deployment-options/) and low-power IoT sensors. +- **Up to 43% Faster CPU Inference:** By removing the Distribution Focal Loss (DFL) module, YOLO26 achieves massive speedups on standard CPUs, making it the premier choice for [mobile deployment](https://docs.ultralytics.com/guides/model-deployment-options) and low-power IoT sensors. - **MuSGD Optimizer:** Inspired by Large Language Model (LLM) training techniques like Moonshot AI's Kimi K2, YOLO26 utilizes a hybrid of SGD and Muon. This ensures incredibly stable training runs and vastly accelerated convergence compared to the AdamW optimizers used in YOLOv10. -- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, which is critical for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and aerial security applications. +- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, which is critical for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and aerial security applications. - **Task-Specific Mastery:** While YOLOv10 is strictly a bounding box detector, YOLO26 offers dedicated architectural improvements for all tasks, including Residual Log-Likelihood Estimation (RLE) for Pose and specialized angle losses for Oriented Bounding Boxes (OBB). !!! note "Explore Further" - If you are exploring the broader landscape of object detection, you may also be interested in comparing these architectures against other frameworks. Check out our deep dives on [YOLO11 vs EfficientDet](https://docs.ultralytics.com/compare/yolo11-vs-efficientdet/) or [RT-DETR vs YOLOv8](https://docs.ultralytics.com/compare/rtdetr-vs-yolov8/) for more comprehensive benchmarking. + If you are exploring the broader landscape of object detection, you may also be interested in comparing these architectures against other frameworks. Check out our deep dives on [YOLO11 vs EfficientDet](https://docs.ultralytics.com/compare/yolo11-vs-efficientdet) or [RT-DETR vs YOLOv8](https://docs.ultralytics.com/compare/rtdetr-vs-yolov8) for more comprehensive benchmarking. Whether you rely on the robust legacy of YOLOv5, the NMS-free innovation of YOLOv10, or the unparalleled cutting-edge performance of YOLO26, the Ultralytics ecosystem provides the tools necessary to bring your vision AI applications to life quickly and efficiently. diff --git a/docs/en/compare/yolov10-vs-yolov6.md b/docs/en/compare/yolov10-vs-yolov6.md index 98aae682493..a61a36e6108 100644 --- a/docs/en/compare/yolov10-vs-yolov6.md +++ b/docs/en/compare/yolov10-vs-yolov6.md @@ -6,7 +6,7 @@ keywords: YOLOv10, YOLOv6, YOLO comparison, object detection models, computer vi # YOLOv10 vs. YOLOv6-3.0: A Comprehensive Technical Comparison -In the rapidly evolving landscape of [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv), selecting the optimal [object detection](https://docs.ultralytics.com/tasks/detect/) architecture is crucial for balancing inference speed, model accuracy, and deployment feasibility. This guide provides an in-depth, technical comparison between two formidable models: the academic powerhouse **YOLOv10** and the industrially focused **YOLOv6-3.0**. Both bring unique architectural innovations to the table, solving distinct challenges in the deployment of real-time vision systems. +In the rapidly evolving landscape of [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv), selecting the optimal [object detection](https://docs.ultralytics.com/tasks/detect) architecture is crucial for balancing inference speed, model accuracy, and deployment feasibility. This guide provides an in-depth, technical comparison between two formidable models: the academic powerhouse **YOLOv10** and the industrially focused **YOLOv6-3.0**. Both bring unique architectural innovations to the table, solving distinct challenges in the deployment of real-time vision systems. @@ -22,13 +22,13 @@ Released in mid-2024, **YOLOv10** introduced a paradigm shift in the YOLO family - **Date:** 2024-05-23 - **ArXiv:** [2405.14458](https://arxiv.org/abs/2405.14458) - **GitHub:** [THU-MIG/yolov10](https://github.com/THU-MIG/yolov10) -- **Docs:** [Ultralytics YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10/) +- **Docs:** [Ultralytics YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10) ### Architectural Innovations -YOLOv10 achieves its NMS-free capability through a **Consistent Dual Assignment** strategy. During training, the model leverages both one-to-many and one-to-one label assignments, enriching supervision signals. For inference, it strictly relies on the one-to-one head, stripping away the computational overhead associated with traditional bounding box filtering. Furthermore, YOLOv10 integrates a holistic, efficiency-driven design, thoroughly optimizing internal components like the [convolutional neural network](https://www.ultralytics.com/glossary/convolutional-neural-network-cnn) layers to drastically reduce computational redundancy and overall [parameter count](https://docs.ultralytics.com/guides/yolo-performance-metrics/). +YOLOv10 achieves its NMS-free capability through a **Consistent Dual Assignment** strategy. During training, the model leverages both one-to-many and one-to-one label assignments, enriching supervision signals. For inference, it strictly relies on the one-to-one head, stripping away the computational overhead associated with traditional bounding box filtering. Furthermore, YOLOv10 integrates a holistic, efficiency-driven design, thoroughly optimizing internal components like the [convolutional neural network](https://www.ultralytics.com/glossary/convolutional-neural-network-cnn) layers to drastically reduce computational redundancy and overall [parameter count](https://docs.ultralytics.com/guides/yolo-performance-metrics). -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## YOLOv6-3.0 Overview: The Industrial Workhorse @@ -39,13 +39,13 @@ Developed specifically for industrial applications, **YOLOv6-3.0** prioritizes h - **Date:** 2023-01-13 - **ArXiv:** [2301.05586](https://arxiv.org/abs/2301.05586) - **GitHub:** [meituan/YOLOv6](https://github.com/meituan/YOLOv6) -- **Docs:** [Ultralytics YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6/) +- **Docs:** [Ultralytics YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6) ### Architectural Innovations -YOLOv6-3.0 distinguishes itself with a heavily optimized **EfficientRep** backbone, structured to maximize inference speeds on hardware accelerators like [NVIDIA GPUs](https://docs.ultralytics.com/guides/nvidia-jetson/). Version 3.0 introduced a **Bi-directional Concatenation (BiC)** module to enhance cross-scale feature fusion. Additionally, it implements an **Anchor-Aided Training (AAT)** strategy that combines the rapid convergence of [anchor-based detectors](https://www.ultralytics.com/glossary/anchor-based-detectors) with the generalization capabilities of anchor-free paradigms. +YOLOv6-3.0 distinguishes itself with a heavily optimized **EfficientRep** backbone, structured to maximize inference speeds on hardware accelerators like [NVIDIA GPUs](https://docs.ultralytics.com/guides/nvidia-jetson). Version 3.0 introduced a **Bi-directional Concatenation (BiC)** module to enhance cross-scale feature fusion. Additionally, it implements an **Anchor-Aided Training (AAT)** strategy that combines the rapid convergence of [anchor-based detectors](https://www.ultralytics.com/glossary/anchor-based-detectors) with the generalization capabilities of anchor-free paradigms. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Performance and Metrics Comparison @@ -65,7 +65,7 @@ When analyzing raw performance, the generations of architectural refinement in Y | YOLOv6-3.0m | 640 | 50.0 | - | 5.28 | 34.9 | 85.8 | | YOLOv6-3.0l | 640 | 52.8 | - | 8.95 | 59.6 | 150.7 | -While YOLOv6-3.0 retains slight speed advantages in its Nano and Medium variants under pure [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) execution on T4 GPUs, YOLOv10 requires nearly half the memory footprint to achieve superior accuracy, heavily leaning the performance balance in favor of modern, end-to-end architectures. +While YOLOv6-3.0 retains slight speed advantages in its Nano and Medium variants under pure [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) execution on T4 GPUs, YOLOv10 requires nearly half the memory footprint to achieve superior accuracy, heavily leaning the performance balance in favor of modern, end-to-end architectures. !!! tip "Memory Efficiency" @@ -76,12 +76,12 @@ While YOLOv6-3.0 retains slight speed advantages in its Nano and Medium variants Opting for an [Ultralytics](https://www.ultralytics.com/) model like YOLOv10 goes far beyond raw architecture—it provides access to a meticulously maintained ecosystem that simplifies the entire machine learning lifecycle. YOLOv6, housed in a static research repository, lacks the robust tooling and multi-task versatility that the Ultralytics framework provides out of the box. - **Ease of Use:** The Ultralytics Python API provides a streamlined user experience, allowing developers to train and export models with just a few lines of code. -- **Versatility:** Unlike YOLOv6, which strictly specializes in detection, the Ultralytics ecosystem empowers you to perform [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) tracking using a unified interface. -- **Well-Maintained Ecosystem:** Enjoy frequent updates, strong community support, and seamless integrations with industry standards like [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) and [ONNX](https://docs.ultralytics.com/integrations/onnx/). +- **Versatility:** Unlike YOLOv6, which strictly specializes in detection, the Ultralytics ecosystem empowers you to perform [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Image Classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) tracking using a unified interface. +- **Well-Maintained Ecosystem:** Enjoy frequent updates, strong community support, and seamless integrations with industry standards like [OpenVINO](https://docs.ultralytics.com/integrations/openvino) and [ONNX](https://docs.ultralytics.com/integrations/onnx). ### Code Example: Consistent Training Workflows -With the Ultralytics SDK, training models is exceptionally straightforward. The system automatically handles complex [data augmentations](https://docs.ultralytics.com/guides/yolo-data-augmentation/) and device scaling. +With the Ultralytics SDK, training models is exceptionally straightforward. The system automatically handles complex [data augmentations](https://docs.ultralytics.com/guides/yolo-data-augmentation) and device scaling. ```python from ultralytics import YOLO @@ -121,11 +121,11 @@ YOLOv6 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultimate Recommendation: Ultralytics YOLO26 diff --git a/docs/en/compare/yolov10-vs-yolov7.md b/docs/en/compare/yolov10-vs-yolov7.md index 54b42d00444..d49f12f99d0 100644 --- a/docs/en/compare/yolov10-vs-yolov7.md +++ b/docs/en/compare/yolov10-vs-yolov7.md @@ -6,14 +6,14 @@ keywords: YOLOv10, YOLOv7, object detection, model comparison, AI, deep learning # YOLOv10 vs YOLOv7: The Evolution of Real-Time Object Detection -The rapid progression of computer vision over the last few years has yielded increasingly efficient architectures for real-time applications. Comparing [YOLOv10](https://docs.ultralytics.com/models/yolov10/) and [YOLOv7](https://docs.ultralytics.com/models/yolov7/) highlights a crucial transition period in this evolution. While YOLOv7 introduced highly effective training strategies and architectural scaling, YOLOv10 revolutionized deployment by eliminating the longstanding reliance on Non-Maximum Suppression (NMS). +The rapid progression of computer vision over the last few years has yielded increasingly efficient architectures for real-time applications. Comparing [YOLOv10](https://docs.ultralytics.com/models/yolov10) and [YOLOv7](https://docs.ultralytics.com/models/yolov7) highlights a crucial transition period in this evolution. While YOLOv7 introduced highly effective training strategies and architectural scaling, YOLOv10 revolutionized deployment by eliminating the longstanding reliance on Non-Maximum Suppression (NMS). -Both models pushed the boundaries of [object detection](https://docs.ultralytics.com/tasks/detect/) upon their respective releases, yet the modern [Ultralytics ecosystem](https://platform.ultralytics.com) and the introduction of next-generation models like YOLO26 offer far superior workflows for today's AI practitioners. +Both models pushed the boundaries of [object detection](https://docs.ultralytics.com/tasks/detect) upon their respective releases, yet the modern [Ultralytics ecosystem](https://platform.ultralytics.com) and the introduction of next-generation models like YOLO26 offer far superior workflows for today's AI practitioners. ## Model Profiles and Origins @@ -26,9 +26,9 @@ Understanding the origins of these models provides valuable context regarding th - Date: 2024-05-23 - Arxiv: [YOLOv10: Real-Time End-to-End Object Detection](https://arxiv.org/abs/2405.14458) - GitHub: [THU-MIG/yolov10](https://github.com/THU-MIG/yolov10) -- Docs: [Ultralytics YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10/) +- Docs: [Ultralytics YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10) -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ### YOLOv7 Details @@ -37,9 +37,9 @@ Understanding the origins of these models provides valuable context regarding th - Date: 2022-07-06 - Arxiv: [YOLOv7: Trainable bag-of-freebies sets new state-of-the-art](https://arxiv.org/abs/2207.02696) - GitHub: [WongKinYiu/yolov7](https://github.com/WongKinYiu/yolov7) -- Docs: [Ultralytics YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7/) +- Docs: [Ultralytics YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7) -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ## Architectural Innovations @@ -49,7 +49,7 @@ Released in 2022, YOLOv7 focused heavily on optimizing gradient pathways. It int ### The YOLOv10 Breakthrough -YOLOv10 addressed the NMS bottleneck directly. By introducing consistent dual assignments during training, the Tsinghua University team enabled NMS-free end-to-end detection. This dual-head approach uses one branch with one-to-many assignments for rich supervisory signals during training, and another branch with one-to-one assignments for NMS-free inference. This architectural shift ensures consistent, ultra-low [inference latency](https://docs.ultralytics.com/guides/yolo-performance-metrics/) suitable for high-speed video analytics. Furthermore, YOLOv10 employs a holistic efficiency-accuracy driven model design, stripping away computational redundancy found in earlier generations. +YOLOv10 addressed the NMS bottleneck directly. By introducing consistent dual assignments during training, the Tsinghua University team enabled NMS-free end-to-end detection. This dual-head approach uses one branch with one-to-many assignments for rich supervisory signals during training, and another branch with one-to-one assignments for NMS-free inference. This architectural shift ensures consistent, ultra-low [inference latency](https://docs.ultralytics.com/guides/yolo-performance-metrics) suitable for high-speed video analytics. Furthermore, YOLOv10 employs a holistic efficiency-accuracy driven model design, stripping away computational redundancy found in earlier generations. !!! tip "Post-Processing Impact" @@ -57,7 +57,7 @@ YOLOv10 addressed the NMS bottleneck directly. By introducing consistent dual as ## Performance Comparison -When comparing raw metrics on the [MS COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/), the generational gap becomes evident. YOLOv10 achieves a much more favorable trade-off between parameters, computational requirements, and accuracy. +When comparing raw metrics on the [MS COCO dataset](https://docs.ultralytics.com/datasets/detect/coco), the generational gap becomes evident. YOLOv10 achieves a much more favorable trade-off between parameters, computational requirements, and accuracy. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | -------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -71,7 +71,7 @@ When comparing raw metrics on the [MS COCO dataset](https://docs.ultralytics.com | YOLOv7l | 640 | 51.4 | - | 6.84 | 36.9 | 104.7 | | YOLOv7x | 640 | 53.1 | - | 11.57 | 71.3 | 189.9 | -As seen above, YOLOv10x delivers a superior mAP of 54.4% compared to YOLOv7x's 53.1%, while using roughly 20% fewer parameters. Furthermore, the lightweight YOLOv10 models (Nano and Small) offer exceptional [TensorRT deployment](https://docs.ultralytics.com/integrations/tensorrt/) speeds, making them highly attractive for mobile deployment. +As seen above, YOLOv10x delivers a superior mAP of 54.4% compared to YOLOv7x's 53.1%, while using roughly 20% fewer parameters. Furthermore, the lightweight YOLOv10 models (Nano and Small) offer exceptional [TensorRT deployment](https://docs.ultralytics.com/integrations/tensorrt) speeds, making them highly attractive for mobile deployment. ## The Ultralytics Ecosystem Advantage @@ -83,11 +83,11 @@ Both YOLOv10 and YOLOv7 can be accessed via the standard Ultralytics Python pack ### Unmatched Versatility -While older repositories often focus strictly on bounding box detection, the integrated Ultralytics framework seamlessly supports a massive variety of tasks. Whether you are performing [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), or [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection, the workflow remains identical. +While older repositories often focus strictly on bounding box detection, the integrated Ultralytics framework seamlessly supports a massive variety of tasks. Whether you are performing [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), or [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection, the workflow remains identical. ### Code Example: Consistent Training Workflows -The following code snippet demonstrates the seamless training process, which automatically handles [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation/) and learning rate scheduling: +The following code snippet demonstrates the seamless training process, which automatically handles [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation) and learning rate scheduling: ```python from ultralytics import YOLO @@ -124,11 +124,11 @@ YOLOv7 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The New Standard: Introducing YOLO26 @@ -163,6 +163,6 @@ YOLOv10 shines in scenarios requiring strict, unchanging latency. Because it is ### When to use YOLO26 -YOLO26 is the definitive choice for any greenfield project. From deploying sophisticated [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/) on a basic Raspberry Pi to running massive cloud-based video analytics, its superior CPU speeds and advanced small-object detection make it vastly superior to older generations. +YOLO26 is the definitive choice for any greenfield project. From deploying sophisticated [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system) on a basic Raspberry Pi to running massive cloud-based video analytics, its superior CPU speeds and advanced small-object detection make it vastly superior to older generations. -For developers interested in exploring alternative modern architectures, we also provide extensive support for transformer-based detectors like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) and previous generational staples like [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11). +For developers interested in exploring alternative modern architectures, we also provide extensive support for transformer-based detectors like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) and previous generational staples like [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11). diff --git a/docs/en/compare/yolov10-vs-yolov8.md b/docs/en/compare/yolov10-vs-yolov8.md index 70bbc497dbb..543aa099336 100644 --- a/docs/en/compare/yolov10-vs-yolov8.md +++ b/docs/en/compare/yolov10-vs-yolov8.md @@ -6,7 +6,7 @@ keywords: YOLOv10, YOLOv8, object detection, model comparison, computer vision, # YOLOv10 vs YOLOv8: A Technical Deep Dive into Modern Object Detection -The evolution of real-time object detection has seen a rapid succession of groundbreaking architectures, each attempting to push the boundaries of accuracy, inference speed, and computational efficiency. In this comprehensive technical guide, we compare two major milestones in the computer vision landscape: **YOLOv10** and **[Ultralytics YOLOv8](https://docs.ultralytics.com/models/yolov8/)**. While YOLOv8 established a highly versatile and production-ready standard, YOLOv10 introduced architectural shifts specifically aimed at removing post-processing bottlenecks. +The evolution of real-time object detection has seen a rapid succession of groundbreaking architectures, each attempting to push the boundaries of accuracy, inference speed, and computational efficiency. In this comprehensive technical guide, we compare two major milestones in the computer vision landscape: **YOLOv10** and **[Ultralytics YOLOv8](https://docs.ultralytics.com/models/yolov8)**. While YOLOv8 established a highly versatile and production-ready standard, YOLOv10 introduced architectural shifts specifically aimed at removing post-processing bottlenecks. Understanding the distinct advantages, architectures, and performance metrics of these models is crucial for developers and researchers aiming to deploy state-of-the-art vision AI solutions in real-world scenarios. @@ -28,9 +28,9 @@ Developed by researchers at Tsinghua University, YOLOv10 was designed to address - **Date:** 2024-05-23 - **Arxiv:** [2405.14458](https://arxiv.org/abs/2405.14458) - **GitHub:** [THU-MIG/yolov10](https://github.com/THU-MIG/yolov10) -- **Docs:** [YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10/) +- **Docs:** [YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10) -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ### Ultralytics YOLOv8: The Versatile Standard @@ -53,7 +53,7 @@ The standout feature of YOLOv10 is its **NMS-free training strategy**. Tradition ### YOLOv8 Architecture -YOLOv8 introduced an **anchor-free detection head**, moving away from the anchor-based approaches of its predecessors. This reduces the number of box predictions and speeds up [NMS operations](https://docs.ultralytics.com/reference/utils/nms/). Additionally, YOLOv8 incorporates the **C2f module** (Cross-Stage Partial bottleneck with two convolutions), which improves gradient flow and allows the network to learn richer feature representations without drastically increasing computational cost. Its decoupled head structure separates objectness, classification, and regression tasks, leading to faster convergence and higher overall accuracy. +YOLOv8 introduced an **anchor-free detection head**, moving away from the anchor-based approaches of its predecessors. This reduces the number of box predictions and speeds up [NMS operations](https://docs.ultralytics.com/reference/utils/nms). Additionally, YOLOv8 incorporates the **C2f module** (Cross-Stage Partial bottleneck with two convolutions), which improves gradient flow and allows the network to learn richer feature representations without drastically increasing computational cost. Its decoupled head structure separates objectness, classification, and regression tasks, leading to faster convergence and higher overall accuracy. ## Performance and Benchmarks @@ -90,7 +90,7 @@ Choosing a model goes beyond theoretical benchmarks; the developer experience an One of the core strengths of YOLOv8 is its tight integration into the [Ultralytics ecosystem](https://docs.ultralytics.com/). This environment provides a "zero-to-hero" experience, characterized by a highly intuitive Python API and extensive documentation. Unlike research-focused repositories that may require complex environment setups, Ultralytics models are renowned for their **ease of use**. -Furthermore, YOLOv8 is inherently versatile. While YOLOv10 is strictly optimized for object detection, the Ultralytics framework allows developers to seamlessly pivot between [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb/) tasks within the exact same library and API structure. +Furthermore, YOLOv8 is inherently versatile. While YOLOv10 is strictly optimized for object detection, the Ultralytics framework allows developers to seamlessly pivot between [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb) tasks within the exact same library and API structure. ### Memory Requirements and Training @@ -132,4 +132,4 @@ When deciding between these architectures, consider the specific needs of your d - **Choose Ultralytics YOLOv8 if:** You need a highly stable, production-ready model supported by the robust [Ultralytics Platform](https://platform.ultralytics.com/). It is the ideal choice if your project requires multiple tasks (e.g., detecting objects and then segmenting them) using a unified, easy-to-maintain codebase. - **Choose YOLO26 (Recommended) if:** You want the ultimate balance of state-of-the-art accuracy, native end-to-end NMS-free efficiency, and the fastest possible speeds on CPU and edge hardware. -If you are exploring the broader landscape, you might also be interested in comparing these models with [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or checking out specific edge-deployment integrations like [Intel OpenVINO](https://docs.ultralytics.com/integrations/openvino/) to further accelerate your vision AI applications. By leveraging the unified tools provided by Ultralytics, deploying robust computer vision solutions has never been more accessible. +If you are exploring the broader landscape, you might also be interested in comparing these models with [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or checking out specific edge-deployment integrations like [Intel OpenVINO](https://docs.ultralytics.com/integrations/openvino) to further accelerate your vision AI applications. By leveraging the unified tools provided by Ultralytics, deploying robust computer vision solutions has never been more accessible. diff --git a/docs/en/compare/yolov10-vs-yolov9.md b/docs/en/compare/yolov10-vs-yolov9.md index d3b461202a1..5a7591cdf25 100644 --- a/docs/en/compare/yolov10-vs-yolov9.md +++ b/docs/en/compare/yolov10-vs-yolov9.md @@ -8,7 +8,7 @@ keywords: YOLOv10,YOLOv9,Ultralytics,object detection,real-time AI,computer visi The evolution of real-time computer vision has been marked by continuous breakthroughs in speed, accuracy, and architectural efficiency. When evaluating modern solutions for your next deployment, comparing **YOLOv10** and **YOLOv9** offers a fascinating look at two distinct approaches to solving deep learning bottlenecks. While YOLOv9 focuses on maximizing gradient information flow during training, YOLOv10 pioneers a native end-to-end design that completely eliminates traditional post-processing hurdles. -This comprehensive guide analyzes their architectural innovations, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/), and ideal use cases to help developers and researchers choose the optimal model for their specific computer vision tasks. +This comprehensive guide analyzes their architectural innovations, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics), and ideal use cases to help developers and researchers choose the optimal model for their specific computer vision tasks. @@ -24,21 +24,21 @@ Developed to address the latency bottlenecks of traditional object detectors, YO - **Authors:** Ao Wang, Hui Chen, Lihao Liu, et al. - **Organization:** Tsinghua University - **Date:** May 23, 2024 -- **Links:** [Arxiv Publication](https://arxiv.org/abs/2405.14458), [GitHub Repository](https://github.com/THU-MIG/yolov10), [Ultralytics Docs](https://docs.ultralytics.com/models/yolov10/) +- **Links:** [Arxiv Publication](https://arxiv.org/abs/2405.14458), [GitHub Repository](https://github.com/THU-MIG/yolov10), [Ultralytics Docs](https://docs.ultralytics.com/models/yolov10) -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ### Architecture and Strengths -YOLOv10's most significant contribution to the field is its consistent dual-assignment strategy for NMS-free training. By eliminating NMS, the model drastically reduces inference latency, especially on edge devices where post-processing can bottleneck the entire pipeline. It optimizes various components from both efficiency and accuracy perspectives, resulting in a model that boasts a remarkable [trade-off between speed and parameters](https://en.wikipedia.org/wiki/Pareto_efficiency). For instance, the YOLOv10-S variant is exceptionally fast, making it highly suitable for high-speed [video analytics](https://docs.ultralytics.com/guides/analytics/) and real-time robotic navigation. +YOLOv10's most significant contribution to the field is its consistent dual-assignment strategy for NMS-free training. By eliminating NMS, the model drastically reduces inference latency, especially on edge devices where post-processing can bottleneck the entire pipeline. It optimizes various components from both efficiency and accuracy perspectives, resulting in a model that boasts a remarkable [trade-off between speed and parameters](https://en.wikipedia.org/wiki/Pareto_efficiency). For instance, the YOLOv10-S variant is exceptionally fast, making it highly suitable for high-speed [video analytics](https://docs.ultralytics.com/guides/analytics) and real-time robotic navigation. ### Weaknesses -While the NMS-free design is groundbreaking for bounding box detection, YOLOv10 is primarily optimized as a pure object detector. It lacks the out-of-the-box versatility of newer ecosystems that natively support [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/) or [Pose Estimation](https://docs.ultralytics.com/tasks/pose/). Furthermore, early implementations required careful export handling to ensure operations like `cv2` were fully optimized out of the inference graph. +While the NMS-free design is groundbreaking for bounding box detection, YOLOv10 is primarily optimized as a pure object detector. It lacks the out-of-the-box versatility of newer ecosystems that natively support [Instance Segmentation](https://docs.ultralytics.com/tasks/segment) or [Pose Estimation](https://docs.ultralytics.com/tasks/pose). Furthermore, early implementations required careful export handling to ensure operations like `cv2` were fully optimized out of the inference graph. !!! tip "Exporting YOLOv10" - When preparing YOLOv10 for production, always ensure you export the model to optimized formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) or ONNX. Running raw PyTorch weights in deployment can result in slower-than-expected inference due to unoptimized graph operations. + When preparing YOLOv10 for production, always ensure you export the model to optimized formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) or ONNX. Running raw PyTorch weights in deployment can result in slower-than-expected inference due to unoptimized graph operations. ## YOLOv9: Programmable Gradient Information @@ -49,9 +49,9 @@ Prior to YOLOv10, YOLOv9 introduced novel architectural concepts to solve the in - **Authors:** Chien-Yao Wang and Hong-Yuan Mark Liao - **Organization:** Institute of Information Science, Academia Sinica, Taiwan - **Date:** February 21, 2024 -- **Links:** [Arxiv Publication](https://arxiv.org/abs/2402.13616), [GitHub Repository](https://github.com/WongKinYiu/yolov9), [Ultralytics Docs](https://docs.ultralytics.com/models/yolov9/) +- **Links:** [Arxiv Publication](https://arxiv.org/abs/2402.13616), [GitHub Repository](https://github.com/WongKinYiu/yolov9), [Ultralytics Docs](https://docs.ultralytics.com/models/yolov9) -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ### Architecture and Strengths @@ -96,7 +96,7 @@ While YOLOv9 and YOLOv10 are impressive milestones, the machine learning landsca !!! note "The Ultralytics Ecosystem Advantage" - Choosing an Ultralytics model like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or YOLO26 provides unparalleled ease of use. You gain access to active development, a thriving community, and frequent updates that ensure your models remain compatible with the latest inference engines like [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) and CoreML. + Choosing an Ultralytics model like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or YOLO26 provides unparalleled ease of use. You gain access to active development, a thriving community, and frequent updates that ensure your models remain compatible with the latest inference engines like [OpenVINO](https://docs.ultralytics.com/integrations/openvino) and CoreML. ## Practical Implementation @@ -140,11 +140,11 @@ YOLOv9 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Conclusion diff --git a/docs/en/compare/yolov10-vs-yolox.md b/docs/en/compare/yolov10-vs-yolox.md index edddd03f4e9..78ac1294e00 100644 --- a/docs/en/compare/yolov10-vs-yolox.md +++ b/docs/en/compare/yolov10-vs-yolox.md @@ -26,9 +26,9 @@ Developed to resolve long-standing latency bottlenecks, YOLOv10 introduced a nat - **Date:** May 23, 2024 - **ArXiv:** [2405.14458](https://arxiv.org/abs/2405.14458) - **GitHub:** [THU-MIG/yolov10](https://github.com/THU-MIG/yolov10) -- **Docs:** [Ultralytics YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10/) +- **Docs:** [Ultralytics YOLOv10 Documentation](https://docs.ultralytics.com/models/yolov10) -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ### YOLOX: Bridging the Research and Industry Gap @@ -51,7 +51,7 @@ Both frameworks diverge from traditional anchor-based detectors, but they solve YOLOX brought several crucial updates to the ecosystem back in 2021. Its primary contribution was shifting to an **anchor-free detector** design. By eliminating predefined anchor boxes, YOLOX heavily reduced the number of design parameters and heuristic tuning required for different datasets. -Furthermore, YOLOX employs a **decoupled head**, separating the classification and regression tasks. This resolved the conflict between the two objectives, significantly accelerating convergence during training. It also utilizes **SimOTA** for advanced label assignment, improving the handling of crowded scenes and occlusions common in the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). +Furthermore, YOLOX employs a **decoupled head**, separating the classification and regression tasks. This resolved the conflict between the two objectives, significantly accelerating convergence during training. It also utilizes **SimOTA** for advanced label assignment, improving the handling of crowded scenes and occlusions common in the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). !!! tip "Anchor-Free Advantage" @@ -83,7 +83,7 @@ Evaluating these models on hardware like the NVIDIA T4 GPU reveals distinct adva | YOLOXl | 640 | 49.7 | - | 9.04 | 54.2 | 155.6 | | YOLOXx | 640 | 51.1 | - | 16.1 | 99.1 | 281.9 | -As seen above, YOLOv10 scales exceptionally well. The `YOLOv10x` variant achieves the highest accuracy (**54.4 mAP**), while the `YOLOv10n` variant delivers the fastest inference using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) integration. Conversely, the legacy YOLOX nano model features the smallest overall footprint for heavily constrained environments. +As seen above, YOLOv10 scales exceptionally well. The `YOLOv10x` variant achieves the highest accuracy (**54.4 mAP**), while the `YOLOv10n` variant delivers the fastest inference using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) integration. Conversely, the legacy YOLOX nano model features the smallest overall footprint for heavily constrained environments. ## Training Methodologies and Resource Requirements @@ -91,7 +91,7 @@ When implementing models for production, the training ecosystem and resource dem YOLOX often relies on older environment configurations that can be cumbersome to manage. Furthermore, its legacy codebase requires more boilerplate code to achieve multi-GPU distributed training or mixed-precision optimization. -In contrast, YOLOv10 integrates smoothly with modern PyTorch workflows, but it is the **Ultralytics ecosystem** that truly transforms the developer experience. Ultralytics models are characterized by significantly lower CUDA memory usage during training compared to transformer-based architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +In contrast, YOLOv10 integrates smoothly with modern PyTorch workflows, but it is the **Ultralytics ecosystem** that truly transforms the developer experience. Ultralytics models are characterized by significantly lower CUDA memory usage during training compared to transformer-based architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). ### Code Example: Streamlined Training @@ -113,7 +113,7 @@ metrics = model.val() model.export(format="onnx") ``` -This simple syntax provides immediate access to [automatic mixed precision](https://www.ultralytics.com/glossary/mixed-precision), automated data augmentation, and integration with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) out of the box. +This simple syntax provides immediate access to [automatic mixed precision](https://www.ultralytics.com/glossary/mixed-precision), automated data augmentation, and integration with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) out of the box. ## Use Cases and Recommendations @@ -137,11 +137,11 @@ YOLOX is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Future of Vision AI: Enter YOLO26 @@ -154,7 +154,7 @@ YOLO26 stands out by introducing several massive leaps forward: - **Up to 43% Faster CPU Inference:** By strategically removing Distribution Focal Loss (DFL), YOLO26 achieves vastly superior performance on edge devices without GPUs. - **MuSGD Optimizer:** Inspired by LLM training stability, this novel hybrid of SGD and Muon ensures faster convergence and highly stable training runs. - **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, a critical factor for aerial imagery and IoT sensors. -- **Unmatched Versatility:** Unlike YOLOX, which is strictly an object detector, YOLO26 natively supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), and [OBB Detection](https://docs.ultralytics.com/tasks/obb/) within a single, unified library. +- **Unmatched Versatility:** Unlike YOLOX, which is strictly an object detector, YOLO26 natively supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Image Classification](https://docs.ultralytics.com/tasks/classify), and [OBB Detection](https://docs.ultralytics.com/tasks/obb) within a single, unified library. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -168,7 +168,7 @@ Choosing the right model dictates the success of real-world deployments across v ### High-Speed Video Analytics -For processing dense video feeds, such as smart city traffic management, **YOLOv10** provides a significant advantage due to its NMS-free post-processing. Eliminating the NMS bottleneck allows for consistent low latency, making it ideal to pair with tracking algorithms like [BoT-SORT](https://docs.ultralytics.com/reference/trackers/bot_sort/). +For processing dense video feeds, such as smart city traffic management, **YOLOv10** provides a significant advantage due to its NMS-free post-processing. Eliminating the NMS bottleneck allows for consistent low latency, making it ideal to pair with tracking algorithms like [BoT-SORT](https://docs.ultralytics.com/reference/trackers/bot_sort). ### Legacy Edge Deployment @@ -178,4 +178,4 @@ For older academic setups or legacy Android applications heavily optimized for p For next-generation hardware deployments, such as robotics, drones, and retail shelf analysis, [YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) is the ultimate solution. Its drastically reduced CPU latency and superior small-object detection make it uniquely qualified for autonomous navigation and granular inventory management. -For additional comparisons to expand your deep learning toolkit, you can also explore how these models stack up against alternatives like the flexible [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or the transformer-powered [RT-DETR](https://docs.ultralytics.com/compare/rtdetr-vs-yolov10/). +For additional comparisons to expand your deep learning toolkit, you can also explore how these models stack up against alternatives like the flexible [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or the transformer-powered [RT-DETR](https://docs.ultralytics.com/compare/rtdetr-vs-yolov10). diff --git a/docs/en/compare/yolov5-vs-damo-yolo.md b/docs/en/compare/yolov5-vs-damo-yolo.md index aa24b00b83a..b162efeda31 100644 --- a/docs/en/compare/yolov5-vs-damo-yolo.md +++ b/docs/en/compare/yolov5-vs-damo-yolo.md @@ -27,7 +27,7 @@ Developed by Glenn Jocher and the team at Ultralytics, YOLOv5 has become an indu - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2020-06-26 - **GitHub:** [https://github.com/ultralytics/yolov5](https://github.com/ultralytics/yolov5) -- **Docs:** [Ultralytics YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5/) +- **Docs:** [Ultralytics YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5) [Learn more about YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5){ .md-button } @@ -51,7 +51,7 @@ Both models leverage unique structural concepts to achieve their real-time perfo YOLOv5 utilizes a Modified CSP (Cross Stage Partial) backbone paired with a PANet (Path Aggregation Network) neck. This structure is highly efficient, minimizing [CUDA](https://developer.nvidia.com/cuda/toolkit) memory usage during both training and inference. -One of YOLOv5's greatest strengths is its [versatility across tasks](https://docs.ultralytics.com/tasks/). Beyond bounding box predictions, it offers dedicated architectures for [image segmentation](https://docs.ultralytics.com/tasks/segment/) and [image classification](https://docs.ultralytics.com/tasks/classify/), allowing developers to standardize their vision pipelines around a single, cohesive framework. +One of YOLOv5's greatest strengths is its [versatility across tasks](https://docs.ultralytics.com/tasks). Beyond bounding box predictions, it offers dedicated architectures for [image segmentation](https://docs.ultralytics.com/tasks/segment) and [image classification](https://docs.ultralytics.com/tasks/classify), allowing developers to standardize their vision pipelines around a single, cohesive framework. ### DAMO-YOLO: Automated Architecture Search @@ -80,7 +80,7 @@ Evaluating real-time object detectors requires looking at a matrix of mAP (mean | DAMO-YOLOm | 640 | 49.2 | - | 5.09 | 28.2 | 61.8 | | DAMO-YOLOl | 640 | **50.8** | - | 7.18 | 42.1 | 97.3 | -While DAMO-YOLO achieves highly competitive mAP scores at certain parameter counts, YOLOv5 consistently demonstrates exceptional [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) speeds and incredibly low parameter counts for its nano and small configurations. This performance balance ensures YOLOv5 operates efficiently across diverse edge deployment scenarios. +While DAMO-YOLO achieves highly competitive mAP scores at certain parameter counts, YOLOv5 consistently demonstrates exceptional [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) speeds and incredibly low parameter counts for its nano and small configurations. This performance balance ensures YOLOv5 operates efficiently across diverse edge deployment scenarios. ## Training Efficiency and Ecosystem @@ -92,7 +92,7 @@ DAMO-YOLO relies heavily on a multi-stage training methodology. It implements a ### The Ultralytics Advantage: Ease of Use -Conversely, the [Ultralytics ecosystem](https://docs.ultralytics.com/) is world-renowned for its intuitive APIs and [training efficiency](https://docs.ultralytics.com/modes/train/). Supported by active development and an enormous open-source community, developers can train, validate, and deploy models seamlessly. +Conversely, the [Ultralytics ecosystem](https://docs.ultralytics.com/) is world-renowned for its intuitive APIs and [training efficiency](https://docs.ultralytics.com/modes/train). Supported by active development and an enormous open-source community, developers can train, validate, and deploy models seamlessly. ```python from ultralytics import YOLO @@ -107,7 +107,7 @@ results = model.train(data="coco8.yaml", epochs=100, imgsz=640) model.export(format="onnx") ``` -Ultralytics also provides built-in support for [experiment tracking](https://docs.ultralytics.com/integrations/weights-biases/) via tools like Weights & Biases and Comet ML, creating a frictionless workflow. +Ultralytics also provides built-in support for [experiment tracking](https://docs.ultralytics.com/integrations/weights-biases) via tools like Weights & Biases and Comet ML, creating a frictionless workflow. ## Real-World Use Cases @@ -124,7 +124,7 @@ YOLOv5 is a strong choice for: - **Proven Production Systems:** Existing deployments where YOLOv5's long track record of stability, extensive documentation, and massive community support are valued. - **Resource-Constrained Training:** Environments with limited GPU resources where YOLOv5's efficient training pipeline and lower memory requirements are advantageous. -- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), [CoreML](https://docs.ultralytics.com/integrations/coreml/), and [TFLite](https://docs.ultralytics.com/integrations/tflite/). +- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), [CoreML](https://docs.ultralytics.com/integrations/coreml), and [TFLite](https://docs.ultralytics.com/integrations/tflite). ### When to Choose DAMO-YOLO @@ -136,11 +136,11 @@ DAMO-YOLO is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Next Evolution: YOLO26 @@ -154,8 +154,8 @@ Key innovations in [YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) - **MuSGD Optimizer:** Inspired by LLM training innovations, this hybrid of SGD and Muon ensures highly stable training and rapid convergence. - **Up to 43% Faster CPU Inference:** Heavily optimized for edge computing, making it perfect for IoT devices operating without dedicated GPUs. -- **ProgLoss + STAL:** Advanced loss functions that drastically improve the recognition of small objects, which is critical for [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and robotics. -- **Task-Specific Improvements:** From specialized angle loss for [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) to Residual Log-Likelihood Estimation (RLE) for accurate [Pose estimation](https://docs.ultralytics.com/tasks/pose/), YOLO26 handles complex domains with ease. +- **ProgLoss + STAL:** Advanced loss functions that drastically improve the recognition of small objects, which is critical for [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and robotics. +- **Task-Specific Improvements:** From specialized angle loss for [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb) to Residual Log-Likelihood Estimation (RLE) for accurate [Pose estimation](https://docs.ultralytics.com/tasks/pose), YOLO26 handles complex domains with ease. ## Conclusion @@ -165,6 +165,6 @@ We highly recommend utilizing the [Ultralytics Platform](https://platform.ultral ### Further Reading -- Explore the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) for high-accuracy applications. +- Explore the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr) for high-accuracy applications. - Learn about the previous generation [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) model. -- Discover how to optimize deployments with [OpenVINO](https://docs.ultralytics.com/integrations/openvino/). +- Discover how to optimize deployments with [OpenVINO](https://docs.ultralytics.com/integrations/openvino). diff --git a/docs/en/compare/yolov5-vs-efficientdet.md b/docs/en/compare/yolov5-vs-efficientdet.md index 0876ff974d6..da316ca5f8f 100644 --- a/docs/en/compare/yolov5-vs-efficientdet.md +++ b/docs/en/compare/yolov5-vs-efficientdet.md @@ -6,7 +6,7 @@ keywords: YOLOv5, EfficientDet, object detection, model comparison, computer vis # YOLOv5 vs. EfficientDet: Evaluating Real-Time Object Detection Architectures -When embarking on a new [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) project, choosing the right neural network architecture is one of the most consequential decisions you will make. This guide provides an in-depth technical comparison between **Ultralytics YOLOv5** and Google's **EfficientDet**. By analyzing their architectures, performance metrics, and training ecosystems, we aim to help developers and researchers identify the best [object detection](https://docs.ultralytics.com/tasks/detect/) model for their specific deployment environments. +When embarking on a new [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) project, choosing the right neural network architecture is one of the most consequential decisions you will make. This guide provides an in-depth technical comparison between **Ultralytics YOLOv5** and Google's **EfficientDet**. By analyzing their architectures, performance metrics, and training ecosystems, we aim to help developers and researchers identify the best [object detection](https://docs.ultralytics.com/tasks/detect) model for their specific deployment environments. While EfficientDet introduced novel concepts in compound scaling and feature fusion, [YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5) revolutionized the industry by democratizing access to high-performance AI through its incredibly intuitive [PyTorch](https://pytorch.org/) implementation, streamlined user experience, and unparalleled balance of speed and accuracy. @@ -63,7 +63,7 @@ EfficientDet's core proposition lies in its systematic approach to scaling and f ## Performance and Metrics Comparison -When comparing these models, evaluating their performance on standard benchmarks like the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/) is crucial. The table below highlights the trade-offs between size, computational demand (FLOPs), and inference speed. +When comparing these models, evaluating their performance on standard benchmarks like the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco) is crucial. The table below highlights the trade-offs between size, computational demand (FLOPs), and inference speed. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | --------------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -84,9 +84,9 @@ When comparing these models, evaluating their performance on standard benchmarks ### Balanced Analysis -**YOLOv5** shines in its deployment flexibility and raw hardware acceleration compatibility. Notice the blisteringly fast TensorRT speeds on the T4 GPU. This makes YOLOv5 incredibly well-suited for high-throughput video analytics and [real-time inference](https://www.ultralytics.com/glossary/real-time-inference) pipelines. Furthermore, the Ultralytics ecosystem makes exporting to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/), [CoreML](https://docs.ultralytics.com/integrations/coreml/), and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) a one-line command. +**YOLOv5** shines in its deployment flexibility and raw hardware acceleration compatibility. Notice the blisteringly fast TensorRT speeds on the T4 GPU. This makes YOLOv5 incredibly well-suited for high-throughput video analytics and [real-time inference](https://www.ultralytics.com/glossary/real-time-inference) pipelines. Furthermore, the Ultralytics ecosystem makes exporting to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx), [CoreML](https://docs.ultralytics.com/integrations/coreml), and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) a one-line command. -**EfficientDet** offers excellent parameter efficiency. For a given parameter count, it often extracts a high [mean Average Precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics/). However, this theoretical efficiency does not always translate to faster wall-clock inference times on edge GPUs due to the complex routing of the BiFPN layer, which can be memory-bandwidth bound rather than compute-bound. +**EfficientDet** offers excellent parameter efficiency. For a given parameter count, it often extracts a high [mean Average Precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics). However, this theoretical efficiency does not always translate to faster wall-clock inference times on edge GPUs due to the complex routing of the BiFPN layer, which can be memory-bandwidth bound rather than compute-bound. ## Ecosystem and Ease of Use @@ -94,7 +94,7 @@ The defining advantage of choosing an Ultralytics model lies in the surrounding With the introduction of the [Ultralytics Platform](https://platform.ultralytics.com), users can seamlessly transition from data collection to deployment. This platform supports auto-annotation, cloud training, and model monitoring out of the box. In contrast, training EfficientDet often requires navigating the complexities of older TensorFlow object detection APIs, which can present a steep learning curve for rapid prototyping. -Furthermore, YOLOv5's versatility extends beyond bounding boxes. Through continuous updates, the Ultralytics framework natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [image classification](https://docs.ultralytics.com/tasks/classify/), providing a unified API for multiple computer vision tasks. +Furthermore, YOLOv5's versatility extends beyond bounding boxes. Through continuous updates, the Ultralytics framework natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [image classification](https://docs.ultralytics.com/tasks/classify), providing a unified API for multiple computer vision tasks. ## Ideal Use Cases @@ -113,7 +113,7 @@ Building upon the legacy of its predecessors (like [YOLOv8](https://platform.ult - **Advanced Loss Functions:** The integration of ProgLoss and STAL drastically improves the recognition of small targets, which is vital for high-altitude drone imagery and [robotics](https://www.ultralytics.com/solutions/ai-in-robotics). - **DFL Removal:** By removing Distribution Focal Loss, the model export process is streamlined, further enhancing compatibility across diverse hardware accelerators. -Users interested in exploring other recent architectures within the Ultralytics ecosystem might also compare models like [YOLOv10](https://docs.ultralytics.com/models/yolov10/) or [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +Users interested in exploring other recent architectures within the Ultralytics ecosystem might also compare models like [YOLOv10](https://docs.ultralytics.com/models/yolov10) or [RT-DETR](https://docs.ultralytics.com/models/rtdetr). !!! tip "Migrating is Easy" diff --git a/docs/en/compare/yolov5-vs-pp-yoloe.md b/docs/en/compare/yolov5-vs-pp-yoloe.md index f57ac0c9554..4cbe386e0a6 100644 --- a/docs/en/compare/yolov5-vs-pp-yoloe.md +++ b/docs/en/compare/yolov5-vs-pp-yoloe.md @@ -27,11 +27,11 @@ Released in mid-2020, [Ultralytics YOLOv5](https://platform.ultralytics.com/ultr - Organization: [Ultralytics](https://www.ultralytics.com) - Date: 2020-06-26 - GitHub: [ultralytics/yolov5](https://github.com/ultralytics/yolov5) -- Docs: [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5/) +- Docs: [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5) YOLOv5 utilizes a modified CSPDarknet backbone, which efficiently captures rich feature representations while maintaining a lightweight parameter count. It introduced auto-learning anchor boxes, automatically calculating the optimal anchor dimensions for custom datasets before training even begins. Furthermore, its integration of mosaic data augmentation significantly enhances the model's ability to detect smaller objects and generalize across complex spatial contexts. -One of the greatest strengths of YOLOv5 is its incredible versatility. Unlike standard object detectors, the YOLOv5 family seamlessly supports [image classification](https://docs.ultralytics.com/tasks/classify/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), and bounding box detection within a unified API. Its highly optimized architecture also translates to substantially lower memory usage during training and inference compared to heavy transformer-based networks. +One of the greatest strengths of YOLOv5 is its incredible versatility. Unlike standard object detectors, the YOLOv5 family seamlessly supports [image classification](https://docs.ultralytics.com/tasks/classify), [instance segmentation](https://docs.ultralytics.com/tasks/segment), and bounding box detection within a unified API. Its highly optimized architecture also translates to substantially lower memory usage during training and inference compared to heavy transformer-based networks. [Learn more about YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5){ .md-button } @@ -54,7 +54,7 @@ PP-YOLOE+ relies on an anchor-free paradigm and utilizes a CSPRepResNet backbone ## Performance and Benchmarks -When evaluating these models for production, understanding the trade-offs between precision, inference speed, and parameter footprint is crucial. The table below outlines key [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/) across different size variants. +When evaluating these models for production, understanding the trade-offs between precision, inference speed, and parameter footprint is crucial. The table below outlines key [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics) across different size variants. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -74,13 +74,13 @@ While PP-YOLOE+ achieves high accuracy limits, YOLOv5 consistently demonstrates !!! tip "Memory Efficiency" - Ultralytics models are specifically engineered for training efficiency. Compared to heavy vision transformers like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/), YOLOv5 uses significantly less CUDA memory, enabling you to train on larger batch sizes or consumer-grade hardware. + Ultralytics models are specifically engineered for training efficiency. Compared to heavy vision transformers like [RT-DETR](https://docs.ultralytics.com/models/rtdetr), YOLOv5 uses significantly less CUDA memory, enabling you to train on larger batch sizes or consumer-grade hardware. ## The Ultralytics Advantage: Ecosystem and Ease of Use The true value of a machine learning architecture extends beyond raw numbers; it encompasses the entire developer experience. The [Ultralytics Platform](https://platform.ultralytics.com) and its corresponding open-source tools provide a highly refined, well-maintained ecosystem that drastically accelerates development cycles. -- **Ease of Use:** Ultralytics abstracts away complex boilerplate code. You can train, validate, and test models via an intuitive [Python API](https://docs.ultralytics.com/usage/python/) or CLI. +- **Ease of Use:** Ultralytics abstracts away complex boilerplate code. You can train, validate, and test models via an intuitive [Python API](https://docs.ultralytics.com/usage/python) or CLI. - **Deployment Flexibility:** Exporting models is incredibly straightforward. With a single command, you can convert your trained YOLOv5 weights to formats like [ONNX](https://onnx.ai/), [TensorRT](https://developer.nvidia.com/tensorrt), or OpenVINO, ensuring broad compatibility across edge and cloud environments. - **Active Community:** The vibrant community guarantees frequent updates, extensive documentation, and robust solutions to common computer vision challenges. @@ -116,7 +116,7 @@ YOLOv5 is a strong choice for: - **Proven Production Systems:** Existing deployments where YOLOv5's long track record of stability, extensive documentation, and massive community support are valued. - **Resource-Constrained Training:** Environments with limited GPU resources where YOLOv5's efficient training pipeline and lower memory requirements are advantageous. -- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), [CoreML](https://docs.ultralytics.com/integrations/coreml/), and [TFLite](https://docs.ultralytics.com/integrations/tflite/). +- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), [CoreML](https://docs.ultralytics.com/integrations/coreml), and [TFLite](https://docs.ultralytics.com/integrations/tflite). ### When to Choose PP-YOLOE+ @@ -128,11 +128,11 @@ PP-YOLOE+ is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Alternative State-of-the-Art Models to Consider @@ -142,10 +142,10 @@ While YOLOv5 is a robust and proven standard, the field of computer vision moves Released in January 2026, [YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) represents the absolute pinnacle of our research. It delivers massive improvements in both accuracy and speed. Key innovations include: -- **End-to-End NMS-Free Design:** Building on concepts from [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing, cutting latency and simplifying deployment logic. +- **End-to-End NMS-Free Design:** Building on concepts from [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing, cutting latency and simplifying deployment logic. - **DFL Removal:** By stripping out Distribution Focal Loss, YOLO26 achieves up to 43% faster CPU inference, making it incredibly powerful for low-power edge devices. - **MuSGD Optimizer:** Inspired by advanced LLM training techniques, this hybrid of SGD and Muon ensures exceptionally stable training runs and faster convergence. -- **ProgLoss + STAL:** These advanced loss functions deliver notable improvements in small-object recognition, which is critical for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and smart agriculture. +- **ProgLoss + STAL:** These advanced loss functions deliver notable improvements in small-object recognition, which is critical for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and smart agriculture. Additionally, you might consider [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), which offers excellent performance and serves as a highly reliable bridge between legacy systems and the bleeding-edge capabilities of YOLO26. @@ -154,7 +154,7 @@ Additionally, you might consider [YOLO11](https://platform.ultralytics.com/ultra The choice between YOLOv5 and PP-YOLOE+ ultimately depends on your deployment environment and project constraints. **Ideal YOLOv5 Applications:** -YOLOv5's minimal resource requirements and incredible ease of use make it the premier choice for [edge AI](https://www.ultralytics.com/glossary/edge-ai). It excels in applications requiring high frame rates on limited hardware, such as real-time [robotics](https://www.ultralytics.com/glossary/robotics), mobile application integration, and multi-camera traffic monitoring systems. Its ability to simultaneously handle [pose estimation](https://docs.ultralytics.com/tasks/pose/) and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb/) tasks within the same framework makes it highly adaptable. +YOLOv5's minimal resource requirements and incredible ease of use make it the premier choice for [edge AI](https://www.ultralytics.com/glossary/edge-ai). It excels in applications requiring high frame rates on limited hardware, such as real-time [robotics](https://www.ultralytics.com/glossary/robotics), mobile application integration, and multi-camera traffic monitoring systems. Its ability to simultaneously handle [pose estimation](https://docs.ultralytics.com/tasks/pose) and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb) tasks within the same framework makes it highly adaptable. **Ideal PP-YOLOE+ Applications:** PP-YOLOE+ is best suited for scenarios where absolute maximum accuracy on static imagery is prioritized over real-time processing constraints. It finds niche usage in industrial inspection pipelines, particularly within Asian manufacturing sectors that have pre-established technical stacks heavily invested in the Baidu and PaddlePaddle ecosystem. diff --git a/docs/en/compare/yolov5-vs-rtdetr.md b/docs/en/compare/yolov5-vs-rtdetr.md index 9679c735b45..c07f16f211b 100644 --- a/docs/en/compare/yolov5-vs-rtdetr.md +++ b/docs/en/compare/yolov5-vs-rtdetr.md @@ -30,9 +30,9 @@ Since its release, Ultralytics YOLOv5 has become a cornerstone of the AI communi YOLOv5 utilizes a streamlined CNN architecture designed to maximize [feature extraction](https://www.ultralytics.com/glossary/feature-extraction) efficiency while maintaining an extremely low memory footprint. It employs a CSPDarknet backbone and a PANet neck, creating a powerful combination for multi-scale feature fusion. -One of the primary advantages of YOLOv5 is its **Performance Balance**. It strikes an exceptional trade-off between speed and accuracy, making it an ideal choice for [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/) on resource-constrained hardware like [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) devices and smartphones. +One of the primary advantages of YOLOv5 is its **Performance Balance**. It strikes an exceptional trade-off between speed and accuracy, making it an ideal choice for [model deployment](https://docs.ultralytics.com/guides/model-deployment-options) on resource-constrained hardware like [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) devices and smartphones. -Furthermore, YOLOv5 boasts unparalleled **Versatility**. Unlike models strictly confined to bounding box predictions, YOLOv5 natively supports [image classification](https://docs.ultralytics.com/tasks/classify/) and [instance segmentation](https://docs.ultralytics.com/tasks/segment/), providing a unified framework for varied visual tasks. Its [training efficiency](https://docs.ultralytics.com/guides/model-training-tips/) is also remarkable, requiring significantly less CUDA memory during training compared to transformer-based architectures. +Furthermore, YOLOv5 boasts unparalleled **Versatility**. Unlike models strictly confined to bounding box predictions, YOLOv5 natively supports [image classification](https://docs.ultralytics.com/tasks/classify) and [instance segmentation](https://docs.ultralytics.com/tasks/segment), providing a unified framework for varied visual tasks. Its [training efficiency](https://docs.ultralytics.com/guides/model-training-tips) is also remarkable, requiring significantly less CUDA memory during training compared to transformer-based architectures. ### Weaknesses @@ -55,13 +55,13 @@ RTDETRv2 (Real-Time Detection Transformer v2) represents a substantial leap in a RTDETRv2 builds upon its predecessor by utilizing a hybrid encoder and a flexible decoder design to process images. The transformer's self-attention mechanism provides the model with a global understanding of the image context, allowing it to perform exceptionally well in complex scenes with severe object occlusion. -A defining feature of RTDETRv2 is its end-to-end, NMS-free design. By predicting object queries directly without requiring [anchor boxes](https://www.ultralytics.com/glossary/anchor-boxes) or NMS post-processing, it simplifies the inference pipeline. This architecture achieves an impressive [mAP (mean Average Precision)](https://www.ultralytics.com/glossary/mean-average-precision-map) on benchmark datasets like [COCO](https://docs.ultralytics.com/datasets/detect/coco/). +A defining feature of RTDETRv2 is its end-to-end, NMS-free design. By predicting object queries directly without requiring [anchor boxes](https://www.ultralytics.com/glossary/anchor-boxes) or NMS post-processing, it simplifies the inference pipeline. This architecture achieves an impressive [mAP (mean Average Precision)](https://www.ultralytics.com/glossary/mean-average-precision-map) on benchmark datasets like [COCO](https://docs.ultralytics.com/datasets/detect/coco). ### Weaknesses -Despite its real-time capabilities, RTDETRv2 has notably higher **memory requirements** compared to YOLO models. The attention mechanisms in transformers scale quadratically with sequence length, which can lead to out-of-memory errors during high-resolution training unless using massive GPU clusters. Additionally, it lacks the out-of-the-box versatility of the Ultralytics ecosystem, primarily focusing only on 2D [object detection](https://docs.ultralytics.com/tasks/detect/) without native support for segmentation or pose estimation. +Despite its real-time capabilities, RTDETRv2 has notably higher **memory requirements** compared to YOLO models. The attention mechanisms in transformers scale quadratically with sequence length, which can lead to out-of-memory errors during high-resolution training unless using massive GPU clusters. Additionally, it lacks the out-of-the-box versatility of the Ultralytics ecosystem, primarily focusing only on 2D [object detection](https://docs.ultralytics.com/tasks/detect) without native support for segmentation or pose estimation. -[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr){ .md-button } ## Performance Comparison Table @@ -90,7 +90,7 @@ When moving a model from a research notebook to a production environment, the so ### Unmatched Ease of Use -Ultralytics models prioritize an incredibly streamlined user experience. Whether you want to train a custom model, run validation, or export to hardware-specific formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) or [ONNX](https://docs.ultralytics.com/integrations/onnx/), the [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) makes it achievable in just a few lines of code. +Ultralytics models prioritize an incredibly streamlined user experience. Whether you want to train a custom model, run validation, or export to hardware-specific formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) or [ONNX](https://docs.ultralytics.com/integrations/onnx), the [Ultralytics Python API](https://docs.ultralytics.com/usage/python) makes it achievable in just a few lines of code. Here is a practical code example demonstrating how simple it is to train and run inference with an Ultralytics model: @@ -110,7 +110,7 @@ inference_results = model.predict("https://ultralytics.com/images/bus.jpg") inference_results[0].show() ``` -This simple, unified API natively supports [experiment tracking](https://docs.ultralytics.com/guides/hyperparameter-tuning/) integrations with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) and [Comet](https://docs.ultralytics.com/integrations/comet/), allowing developers to log metrics seamlessly without writing complex boilerplate code. +This simple, unified API natively supports [experiment tracking](https://docs.ultralytics.com/guides/hyperparameter-tuning) integrations with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) and [Comet](https://docs.ultralytics.com/integrations/comet), allowing developers to log metrics seamlessly without writing complex boilerplate code. ## Use Cases and Recommendations @@ -122,7 +122,7 @@ YOLOv5 is a strong choice for: - **Proven Production Systems:** Existing deployments where YOLOv5's long track record of stability, extensive documentation, and massive community support are valued. - **Resource-Constrained Training:** Environments with limited GPU resources where YOLOv5's efficient training pipeline and lower memory requirements are advantageous. -- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), [CoreML](https://docs.ultralytics.com/integrations/coreml/), and [TFLite](https://docs.ultralytics.com/integrations/tflite/). +- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), [CoreML](https://docs.ultralytics.com/integrations/coreml), and [TFLite](https://docs.ultralytics.com/integrations/tflite). ### When to Choose RT-DETR @@ -134,19 +134,19 @@ RT-DETR is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Looking Forward: YOLO11 and YOLO26 If you are starting a new vision project today, it is highly recommended to explore the latest generations of Ultralytics models. -While YOLOv5 remains incredibly reliable, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) offers improved accuracy and an expanded set of tasks including [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. +While YOLOv5 remains incredibly reliable, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) offers improved accuracy and an expanded set of tasks including [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. -Even more significantly, the cutting-edge [YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) merges the best of both worlds. It implements an **End-to-End NMS-Free Design** (first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/)), eliminating the post-processing overhead while maintaining the efficiency of a CNN. YOLO26 also introduces the **MuSGD Optimizer**, inspired by LLM training innovations, for faster convergence. With **DFL Removal** (Distribution Focal Loss removed for simplified export and better edge/low-power device compatibility), YOLO26 delivers **Up to 43% Faster CPU Inference**, making it the absolute best choice for edge AI. Additionally, **ProgLoss + STAL** provides improved loss functions with notable improvements in small-object recognition, critical for IoT, robotics, and aerial imagery. +Even more significantly, the cutting-edge [YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) merges the best of both worlds. It implements an **End-to-End NMS-Free Design** (first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10)), eliminating the post-processing overhead while maintaining the efficiency of a CNN. YOLO26 also introduces the **MuSGD Optimizer**, inspired by LLM training innovations, for faster convergence. With **DFL Removal** (Distribution Focal Loss removed for simplified export and better edge/low-power device compatibility), YOLO26 delivers **Up to 43% Faster CPU Inference**, making it the absolute best choice for edge AI. Additionally, **ProgLoss + STAL** provides improved loss functions with notable improvements in small-object recognition, critical for IoT, robotics, and aerial imagery. ## Conclusion diff --git a/docs/en/compare/yolov5-vs-yolo11.md b/docs/en/compare/yolov5-vs-yolo11.md index 8807ff792b7..d3765ae0901 100644 --- a/docs/en/compare/yolov5-vs-yolo11.md +++ b/docs/en/compare/yolov5-vs-yolo11.md @@ -25,7 +25,7 @@ Released in the summer of 2020, YOLOv5 quickly became an industry standard due t - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2020-06-26 - **GitHub:** [ultralytics/yolov5](https://github.com/ultralytics/yolov5) -- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5/) +- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5) YOLOv5 established a strong baseline for ease of use and introduced powerful training methodologies, including advanced mosaic data augmentation and auto-anchoring. It remains incredibly popular for researchers building upon a well-documented, heavily tested codebase. @@ -39,9 +39,9 @@ Building upon years of feedback and architectural research, YOLO11 was introduce - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2024-09-27 - **GitHub:** [ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11/) +- **Docs:** [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11) -YOLO11 offers a streamlined user experience through the `ultralytics` Python package, boasting a simple API that unifies [object detection](https://docs.ultralytics.com/tasks/detect/), instance segmentation, classification, pose estimation, and oriented bounding boxes (OBB). It achieves a highly favorable trade-off between speed and accuracy, making it ideal for diverse real-world deployment scenarios. +YOLO11 offers a streamlined user experience through the `ultralytics` Python package, boasting a simple API that unifies [object detection](https://docs.ultralytics.com/tasks/detect), instance segmentation, classification, pose estimation, and oriented bounding boxes (OBB). It achieves a highly favorable trade-off between speed and accuracy, making it ideal for diverse real-world deployment scenarios. [Learn more about YOLO11](https://platform.ultralytics.com/ultralytics/yolo11){ .md-button } @@ -75,7 +75,7 @@ The metrics highlight a clear leap in the **performance balance** achieved by YO The performance improvements in YOLO11 stem from several key architectural evolutions. While YOLOv5 utilized a standard CSPNet backbone with C3 modules, YOLO11 introduced more efficient feature extraction blocks like C2f and later C3k2, which optimize gradient flow and reduce computational overhead. -YOLO11 also features a heavily refined head. Moving away from the anchor-based design of older models, newer Ultralytics architectures adopt an anchor-free approach. This reduces the number of box predictions, streamlining the post-processing pipeline and improving the model's ability to generalize across different scales and aspect ratios. Additionally, these models boast superior [training efficiency](https://docs.ultralytics.com/guides/model-training-tips/) and readily available pre-trained weights that accelerate the convergence of fine-tuned datasets. +YOLO11 also features a heavily refined head. Moving away from the anchor-based design of older models, newer Ultralytics architectures adopt an anchor-free approach. This reduces the number of box predictions, streamlining the post-processing pipeline and improving the model's ability to generalize across different scales and aspect ratios. Additionally, these models boast superior [training efficiency](https://docs.ultralytics.com/guides/model-training-tips) and readily available pre-trained weights that accelerate the convergence of fine-tuned datasets. ## Implementation and Code Examples @@ -121,7 +121,7 @@ results.print() !!! info "Deployment Flexibility" - Both models support extensive export formats. Whether you are targeting an [NVIDIA Jetson](https://developer.nvidia.com/embedded-computing) using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) or an iOS application using CoreML, the deployment process is thoroughly documented and supported by the community. + Both models support extensive export formats. Whether you are targeting an [NVIDIA Jetson](https://developer.nvidia.com/embedded-computing) using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) or an iOS application using CoreML, the deployment process is thoroughly documented and supported by the community. ## Ideal Use Cases @@ -129,12 +129,12 @@ Choosing between these models depends largely on your project's lifecycle stage ### When to Choose YOLOv5 -- **Maintaining Legacy Codebases:** If your production environment is heavily customized around the YOLOv5 repository structure or specific [hyperparameter evolution](https://docs.ultralytics.com/yolov5/tutorials/hyperparameter_evolution/) techniques. +- **Maintaining Legacy Codebases:** If your production environment is heavily customized around the YOLOv5 repository structure or specific [hyperparameter evolution](https://docs.ultralytics.com/yolov5/tutorials/hyperparameter_evolution) techniques. - **Academic Baselines:** When publishing research that requires direct benchmarking against established 2020-2022 computer vision standards. ### When to Choose YOLO11 -- **Multi-Task Projects:** When your application requires a mix of tasks such as [pose estimation](https://docs.ultralytics.com/tasks/pose/) and [instance segmentation](https://docs.ultralytics.com/tasks/segment/) using a single, unified API. +- **Multi-Task Projects:** When your application requires a mix of tasks such as [pose estimation](https://docs.ultralytics.com/tasks/pose) and [instance segmentation](https://docs.ultralytics.com/tasks/segment) using a single, unified API. - **Edge Deployments:** For [edge computing](https://www.ultralytics.com/glossary/edge-computing) scenarios where squeezing out maximum mAP for a given computational budget (FLOPs) is critical. - **Commercial AI Solutions:** Ideal for enterprise applications in retail and security, leveraging the robust support of the [Ultralytics Platform](https://platform.ultralytics.com/). @@ -148,8 +148,8 @@ Released in January 2026, YOLO26 introduces paradigm-shifting advancements desig - **MuSGD Optimizer:** Inspired by LLM training innovations from models like Moonshot AI's Kimi K2, this hybrid of SGD and Muon ensures incredibly stable training and dramatically faster convergence. - **Unprecedented CPU Speed:** By removing Distribution Focal Loss (DFL), YOLO26 achieves up to **43% faster CPU inference**, making it the absolute best choice for edge devices and environments without dedicated GPUs. - **Advanced Loss Functions:** The integration of ProgLoss and STAL yields notable improvements in small-object recognition, which is critical for drone analytics, IoT, and robotics. -- **Task-Specific Enhancements:** It introduces specialized optimizations, such as Residual Log-Likelihood Estimation (RLE) for Pose and specialized angle loss for [oriented bounding boxes](https://docs.ultralytics.com/tasks/obb/), ensuring superior performance across all computer vision tasks. +- **Task-Specific Enhancements:** It introduces specialized optimizations, such as Residual Log-Likelihood Estimation (RLE) for Pose and specialized angle loss for [oriented bounding boxes](https://docs.ultralytics.com/tasks/obb), ensuring superior performance across all computer vision tasks. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } -For users interested in specialized architectures beyond standard object detection, you might also explore models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) for transformer-based detection, or [YOLO-World](https://docs.ultralytics.com/models/yolo-world/) for open-vocabulary tracking and detection. Embracing these well-maintained, highly optimized tools ensures your computer vision pipelines remain efficient, scalable, and ahead of the curve. +For users interested in specialized architectures beyond standard object detection, you might also explore models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) for transformer-based detection, or [YOLO-World](https://docs.ultralytics.com/models/yolo-world) for open-vocabulary tracking and detection. Embracing these well-maintained, highly optimized tools ensures your computer vision pipelines remain efficient, scalable, and ahead of the curve. diff --git a/docs/en/compare/yolov5-vs-yolo26.md b/docs/en/compare/yolov5-vs-yolo26.md index 37538fbd28b..9c0192d9aae 100644 --- a/docs/en/compare/yolov5-vs-yolo26.md +++ b/docs/en/compare/yolov5-vs-yolo26.md @@ -25,7 +25,7 @@ Released in 2020, YOLOv5 revolutionized the accessibility of object detection. B - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2020-06-26 - **GitHub:** [https://github.com/ultralytics/yolov5](https://github.com/ultralytics/yolov5) -- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5/) +- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5) YOLOv5 established the foundation for the highly maintained Ultralytics ecosystem. It introduced aggressive data augmentation techniques, efficient training loops, and highly optimized export paths to edge formats like [CoreML](https://developer.apple.com/documentation/coreml) and [ONNX](https://onnx.ai/). Its ease of use and low memory requirements during training made it a staple for startups and researchers worldwide. @@ -39,7 +39,7 @@ Fast forward to January 2026, **Ultralytics YOLO26** represents the pinnacle of - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2026-01-14 - **GitHub:** [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26/) +- **Docs:** [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26) YOLO26 sets a new benchmark for performance balance, offering state-of-the-art accuracy while being explicitly engineered to dominate edge computing scenarios. @@ -47,7 +47,7 @@ YOLO26 sets a new benchmark for performance balance, offering state-of-the-art a !!! tip "Other Ultralytics Models" - If you are migrating an older codebase, you might also be interested in comparing YOLOv5 with [YOLO11](https://docs.ultralytics.com/models/yolo11/), the previous generation model that introduced initial support for diverse tasks like Pose Estimation and Oriented Bounding Boxes (OBB). + If you are migrating an older codebase, you might also be interested in comparing YOLOv5 with [YOLO11](https://docs.ultralytics.com/models/yolo11), the previous generation model that introduced initial support for diverse tasks like Pose Estimation and Oriented Bounding Boxes (OBB). ## Architectural Breakthroughs in YOLO26 @@ -80,13 +80,13 @@ _Note: The YOLO26 Nano (YOLO26n) achieves a staggering 40.9 mAP compared to YOLO ## Versatility and Task Support -YOLOv5 is primarily renowned for [object detection](https://docs.ultralytics.com/tasks/detect/). While later updates introduced basic segmentation, YOLO26 was built from the ground up to be a unified multi-task engine. +YOLOv5 is primarily renowned for [object detection](https://docs.ultralytics.com/tasks/detect). While later updates introduced basic segmentation, YOLO26 was built from the ground up to be a unified multi-task engine. YOLO26 inherently supports: - **Instance Segmentation:** Featuring task-specific multi-scale protos and semantic segmentation loss. - **Pose Estimation:** Utilizing Residual Log-Likelihood Estimation (RLE) for highly accurate keypoint detection. -- **Oriented Bounding Boxes (OBB):** Including specialized angle loss to resolve boundary discontinuity issues, critical for [satellite image analysis](https://docs.ultralytics.com/datasets/obb/dota-v2/). +- **Oriented Bounding Boxes (OBB):** Including specialized angle loss to resolve boundary discontinuity issues, critical for [satellite image analysis](https://docs.ultralytics.com/datasets/obb/dota-v2). - **Image Classification:** Standard full-image categorization. !!! info "Ecosystem Integration" @@ -139,9 +139,9 @@ yolo predict model=yolo26n.engine source=path/to/video.mp4 For any modern computer vision project, **YOLO26 is the undisputed recommendation**. -- **Edge AI and IoT:** Its 43% faster CPU inference and removal of DFL make it perfect for deployment on a [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or mobile devices. -- **High-Speed Pipelines:** The NMS-free architecture ensures stable, predictable latency which is crucial for autonomous robotics and real-time [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/). -- **Complex Scenarios:** If your application requires tracking small objects (e.g., [drone monitoring](https://docs.ultralytics.com/datasets/detect/visdrone/)) or rotating objects (OBB), YOLO26's advanced loss functions (ProgLoss + STAL) provide a massive accuracy advantage. +- **Edge AI and IoT:** Its 43% faster CPU inference and removal of DFL make it perfect for deployment on a [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or mobile devices. +- **High-Speed Pipelines:** The NMS-free architecture ensures stable, predictable latency which is crucial for autonomous robotics and real-time [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system). +- **Complex Scenarios:** If your application requires tracking small objects (e.g., [drone monitoring](https://docs.ultralytics.com/datasets/detect/visdrone)) or rotating objects (OBB), YOLO26's advanced loss functions (ProgLoss + STAL) provide a massive accuracy advantage. ### When to choose YOLOv5 diff --git a/docs/en/compare/yolov5-vs-yolov10.md b/docs/en/compare/yolov5-vs-yolov10.md index 54be17e4e26..e3aadcff24f 100644 --- a/docs/en/compare/yolov5-vs-yolov10.md +++ b/docs/en/compare/yolov5-vs-yolov10.md @@ -6,7 +6,7 @@ keywords: YOLOv5, YOLOv10, object detection, Ultralytics, machine learning model # YOLOv5 vs. YOLOv10: A Comprehensive Technical Comparison -The field of real-time computer vision has seen exponential growth over the past few years, with various architectures pushing the boundaries of what is possible on modern hardware. When evaluating state-of-the-art architectures, the comparison between [YOLOv5](https://docs.ultralytics.com/models/yolov5/) and [YOLOv10](https://docs.ultralytics.com/models/yolov10/) highlights a significant evolutionary step in the domain of object detection. This technical deep dive explores their architectural paradigms, performance trade-offs, and how developers can leverage these tools in production environments. +The field of real-time computer vision has seen exponential growth over the past few years, with various architectures pushing the boundaries of what is possible on modern hardware. When evaluating state-of-the-art architectures, the comparison between [YOLOv5](https://docs.ultralytics.com/models/yolov5) and [YOLOv10](https://docs.ultralytics.com/models/yolov10) highlights a significant evolutionary step in the domain of object detection. This technical deep dive explores their architectural paradigms, performance trade-offs, and how developers can leverage these tools in production environments. @@ -25,11 +25,11 @@ Introduced by Ultralytics, YOLOv5 has long been recognized for its unmatched bal - Organization: [Ultralytics](https://www.ultralytics.com/) - Date: 2020-06-26 - GitHub: [YOLOv5 Repository](https://github.com/ultralytics/yolov5) -- Documentation: [YOLOv5 Docs](https://docs.ultralytics.com/models/yolov5/) +- Documentation: [YOLOv5 Docs](https://docs.ultralytics.com/models/yolov5) [Learn more about YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5){ .md-button } -YOLOv5 relies on an anchor-based detection mechanism combined with a deeply optimized CSPDarknet backbone. This architecture relies heavily on standard operations supported across virtually all inference engines, making it incredibly versatile. Its major strength lies in the [Ultralytics Python SDK](https://docs.ultralytics.com/usage/python/), which provides a streamlined user experience, a simple API, and extensive documentation. Additionally, YOLOv5's lower memory requirements compared to transformer-based models mean it trains rapidly on consumer-grade GPUs without the steep VRAM overhead. +YOLOv5 relies on an anchor-based detection mechanism combined with a deeply optimized CSPDarknet backbone. This architecture relies heavily on standard operations supported across virtually all inference engines, making it incredibly versatile. Its major strength lies in the [Ultralytics Python SDK](https://docs.ultralytics.com/usage/python), which provides a streamlined user experience, a simple API, and extensive documentation. Additionally, YOLOv5's lower memory requirements compared to transformer-based models mean it trains rapidly on consumer-grade GPUs without the steep VRAM overhead. ### YOLOv10: Advancing the Paradigm @@ -40,9 +40,9 @@ Developed by researchers at Tsinghua University, YOLOv10 aimed to address specif - Date: 2024-05-23 - ArXiv: [2405.14458](https://arxiv.org/abs/2405.14458) - GitHub: [YOLOv10 Repository](https://github.com/THU-MIG/yolov10) -- Documentation: [YOLOv10 Docs](https://docs.ultralytics.com/models/yolov10/) +- Documentation: [YOLOv10 Docs](https://docs.ultralytics.com/models/yolov10) -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } The defining characteristic of YOLOv10 is its natively NMS-free (Non-Maximum Suppression) design. By using consistent dual assignments during training, the model eliminates the need for NMS post-processing during inference. This theoretical latency reduction is highly beneficial for deployments running on high-end hardware with powerful [NVIDIA TensorRT](https://developer.nvidia.com/tensorrt) acceleration, though it can introduce structural complexities for edge devices. @@ -69,11 +69,11 @@ When comparing these models, the balance between accuracy (mAP) and computationa | YOLOv10l | 640 | 53.3 | - | 8.33 | 29.5 | 120.3 | | YOLOv10x | 640 | **54.4** | - | 12.2 | 56.9 | 160.4 | -YOLOv10 clearly achieves a higher `mAP50-95` at equivalent size scales, leveraging its modernized efficiency-accuracy driven model design. However, YOLOv5 maintains incredibly competitive latency, especially at the Nano and Small tiers, making it highly reliable for constrained embedded environments like the [NVIDIA Jetson](https://developer.nvidia.com/embedded-computing) line or standard CPUs via [OpenVINO](https://docs.ultralytics.com/integrations/openvino/). +YOLOv10 clearly achieves a higher `mAP50-95` at equivalent size scales, leveraging its modernized efficiency-accuracy driven model design. However, YOLOv5 maintains incredibly competitive latency, especially at the Nano and Small tiers, making it highly reliable for constrained embedded environments like the [NVIDIA Jetson](https://developer.nvidia.com/embedded-computing) line or standard CPUs via [OpenVINO](https://docs.ultralytics.com/integrations/openvino). ## Training Methodologies and Ecosystem -A model's value is deeply tied to the ecosystem surrounding it. Ultralytics maintains an exceptionally well-maintained ecosystem that supports an incredibly wide array of tasks. While YOLOv10 focuses strictly on 2D [object detection](https://docs.ultralytics.com/tasks/detect/), Ultralytics natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +A model's value is deeply tied to the ecosystem surrounding it. Ultralytics maintains an exceptionally well-maintained ecosystem that supports an incredibly wide array of tasks. While YOLOv10 focuses strictly on 2D [object detection](https://docs.ultralytics.com/tasks/detect), Ultralytics natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb). Furthermore, training an Ultralytics model requires significantly lower memory overhead than competing transformer-based methods, keeping the development cycle fast and cost-effective. @@ -113,7 +113,7 @@ YOLOv5 is a strong choice for: - **Proven Production Systems:** Existing deployments where YOLOv5's long track record of stability, extensive documentation, and massive community support are valued. - **Resource-Constrained Training:** Environments with limited GPU resources where YOLOv5's efficient training pipeline and lower memory requirements are advantageous. -- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), [CoreML](https://docs.ultralytics.com/integrations/coreml/), and [TFLite](https://docs.ultralytics.com/integrations/tflite/). +- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), [CoreML](https://docs.ultralytics.com/integrations/coreml), and [TFLite](https://docs.ultralytics.com/integrations/tflite). ### When to Choose YOLOv10 @@ -125,11 +125,11 @@ YOLOv10 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Future: Ultralytics YOLO26 @@ -148,4 +148,4 @@ You can manage, train, and deploy YOLO26 directly via the [Ultralytics Platform] Choosing between YOLOv5 and YOLOv10 often comes down to specific project constraints. YOLOv10 offers excellent mAP for researchers and applications leveraging raw GPU throughput. Conversely, YOLOv5 remains a steadfast, highly compatible workhorse for standard deployments. -However, the field of computer vision is dynamic. To harness the absolute best performance balance, versatility, and ease of use, developers should look to [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/). It encapsulates the speed of NMS-free inference with the robust, well-documented Ultralytics ecosystem, ensuring your vision AI solutions are future-proof. For specialized use cases, developers may also explore [YOLO11](https://docs.ultralytics.com/models/yolo11/) for general robustness, or [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) for transformer-based precision. +However, the field of computer vision is dynamic. To harness the absolute best performance balance, versatility, and ease of use, developers should look to [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26). It encapsulates the speed of NMS-free inference with the robust, well-documented Ultralytics ecosystem, ensuring your vision AI solutions are future-proof. For specialized use cases, developers may also explore [YOLO11](https://docs.ultralytics.com/models/yolo11) for general robustness, or [RT-DETR](https://docs.ultralytics.com/models/rtdetr) for transformer-based precision. diff --git a/docs/en/compare/yolov5-vs-yolov6.md b/docs/en/compare/yolov5-vs-yolov6.md index d560dd1dff1..96939735ae6 100644 --- a/docs/en/compare/yolov5-vs-yolov6.md +++ b/docs/en/compare/yolov5-vs-yolov6.md @@ -40,7 +40,7 @@ Developed by the Vision AI Department at Meituan, YOLOv6-3.0 is tailored specifi YOLOv6 aims to maximize processing speed on GPUs like the [NVIDIA T4](https://www.nvidia.com/en-us/data-center/tesla-t4/). It uses custom quantization methods and specialized backbones to achieve its performance, making it a strong candidate for backend server processing where batch inference is heavily utilized. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Architectural Differences @@ -54,7 +54,7 @@ YOLOv5 utilizes a highly optimized CSPDarknet backbone combined with a Path Aggr Ultralytics models are specifically engineered for training efficiency. You can often train a YOLOv5 model on a single mid-range GPU, making it highly accessible for researchers and startups alike. -Furthermore, YOLOv5 is not just an object detector. Its architecture seamlessly extends to other tasks, offering robust out-of-the-box support for [image segmentation](https://docs.ultralytics.com/tasks/segment/) and [image classification](https://docs.ultralytics.com/tasks/classify/). +Furthermore, YOLOv5 is not just an object detector. Its architecture seamlessly extends to other tasks, offering robust out-of-the-box support for [image segmentation](https://docs.ultralytics.com/tasks/segment) and [image classification](https://docs.ultralytics.com/tasks/classify). ### The YOLOv6-3.0 Architecture @@ -64,7 +64,7 @@ During training, YOLOv6 uses an Anchor-Aided Training (AAT) strategy to stabiliz ## Performance Analysis -When evaluating these models, raw speed and accuracy metrics are vital. Below is a comparative table highlighting the performance of various model sizes on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). +When evaluating these models, raw speed and accuracy metrics are vital. Below is a comparative table highlighting the performance of various model sizes on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ----------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -85,7 +85,7 @@ While YOLOv6-3.0 achieves higher mAP scores in its larger variants, YOLOv5 maint The true defining factor for many engineering teams is the ecosystem surrounding the model. -YOLOv6 is an impressive research repository, but it requires substantial boilerplate code to deploy across varying formats. In contrast, Ultralytics offers a well-maintained ecosystem characterized by a streamlined user experience. Through the unified Python API and the intuitive [Ultralytics Platform](https://platform.ultralytics.com), developers gain access to seamless dataset management, one-click training, and direct exports to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). +YOLOv6 is an impressive research repository, but it requires substantial boilerplate code to deploy across varying formats. In contrast, Ultralytics offers a well-maintained ecosystem characterized by a streamlined user experience. Through the unified Python API and the intuitive [Ultralytics Platform](https://platform.ultralytics.com), developers gain access to seamless dataset management, one-click training, and direct exports to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). ### Code Example: Unified Ultralytics API @@ -117,7 +117,7 @@ YOLOv5 is a strong choice for: - **Proven Production Systems:** Existing deployments where YOLOv5's long track record of stability, extensive documentation, and massive community support are valued. - **Resource-Constrained Training:** Environments with limited GPU resources where YOLOv5's efficient training pipeline and lower memory requirements are advantageous. -- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), [CoreML](https://docs.ultralytics.com/integrations/coreml/), and [TFLite](https://docs.ultralytics.com/integrations/tflite/). +- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), [CoreML](https://docs.ultralytics.com/integrations/coreml), and [TFLite](https://docs.ultralytics.com/integrations/tflite). ### When to Choose YOLOv6 @@ -129,11 +129,11 @@ YOLOv6 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Moving Forward: The YOLO26 Advantage @@ -144,9 +144,9 @@ Released in January 2026, [YOLO26](https://platform.ultralytics.com/ultralytics/ - **End-to-End NMS-Free Design:** YOLO26 eliminates Non-Maximum Suppression post-processing, dramatically reducing latency variance and simplifying deployment logic. - **Up to 43% Faster CPU Inference:** With DFL removal and an optimized head, it drastically outperforms previous generations on edge and low-power devices. - **MuSGD Optimizer:** Leveraging LLM training innovations, the new MuSGD optimizer ensures highly stable training and remarkably fast convergence. -- **Advanced Versatility:** YOLO26 seamlessly handles [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), and Segmentation with specialized task losses like ProgLoss and STAL for unparalleled small-object recognition. +- **Advanced Versatility:** YOLO26 seamlessly handles [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), and Segmentation with specialized task losses like ProgLoss and STAL for unparalleled small-object recognition. -If you are exploring other options within the Ultralytics ecosystem, you might also consider the general-purpose [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or the innovative [YOLO-World](https://docs.ultralytics.com/models/yolo-world/) for open-vocabulary detection tasks. +If you are exploring other options within the Ultralytics ecosystem, you might also consider the general-purpose [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or the innovative [YOLO-World](https://docs.ultralytics.com/models/yolo-world) for open-vocabulary detection tasks. ## Conclusion diff --git a/docs/en/compare/yolov5-vs-yolov7.md b/docs/en/compare/yolov5-vs-yolov7.md index b6e1fc7a4c2..8cff2600340 100644 --- a/docs/en/compare/yolov5-vs-yolov7.md +++ b/docs/en/compare/yolov5-vs-yolov7.md @@ -23,7 +23,7 @@ Understanding the origins and design philosophies behind these models provides c - **Organization:** [Ultralytics](https://www.ultralytics.com) - **Date:** 2020-06-26 - **GitHub:** [YOLOv5 Repository](https://github.com/ultralytics/yolov5) -- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5/) +- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5) [Learn more about YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5){ .md-button } @@ -34,13 +34,13 @@ Understanding the origins and design philosophies behind these models provides c - **Date:** 2022-07-06 - **Arxiv:** [YOLOv7 Paper](https://arxiv.org/abs/2207.02696) - **GitHub:** [YOLOv7 Repository](https://github.com/WongKinYiu/yolov7) -- **Docs:** [YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7/) +- **Docs:** [YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7) -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } !!! tip "Explore More Architectures" - Interested in how these models stack up against others? Check out our comparisons like [YOLOv5 vs YOLO11](https://docs.ultralytics.com/compare/yolo11-vs-yolov5/) or [YOLOv7 vs EfficientDet](https://docs.ultralytics.com/compare/yolov7-vs-efficientdet/) to expand your understanding of the object detection ecosystem. + Interested in how these models stack up against others? Check out our comparisons like [YOLOv5 vs YOLO11](https://docs.ultralytics.com/compare/yolo11-vs-yolov5) or [YOLOv7 vs EfficientDet](https://docs.ultralytics.com/compare/yolov7-vs-efficientdet) to expand your understanding of the object detection ecosystem. ## Architectural Innovations and Differences @@ -48,7 +48,7 @@ Understanding the origins and design philosophies behind these models provides c Introduced by Ultralytics in 2020, YOLOv5 brought a paradigm shift by natively utilizing the [PyTorch](https://pytorch.org/) framework, significantly lowering the barrier to entry for researchers and developers. Its architecture relies on a Modified CSPDarknet53 backbone, integrating Cross Stage Partial (CSP) networks to reduce parameter count while maintaining gradient flow. -One of its greatest strengths is its **Memory requirements**. Compared to older two-stage detectors or heavy transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/), YOLOv5 requires substantially less CUDA memory during training, allowing for larger batch sizes on standard consumer-grade GPUs. Furthermore, its natively integrated **Versatility** supports [image classification](https://docs.ultralytics.com/tasks/classify/), [object detection](https://docs.ultralytics.com/tasks/detect/), and [image segmentation](https://docs.ultralytics.com/tasks/segment/) seamlessly. +One of its greatest strengths is its **Memory requirements**. Compared to older two-stage detectors or heavy transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr), YOLOv5 requires substantially less CUDA memory during training, allowing for larger batch sizes on standard consumer-grade GPUs. Furthermore, its natively integrated **Versatility** supports [image classification](https://docs.ultralytics.com/tasks/classify), [object detection](https://docs.ultralytics.com/tasks/detect), and [image segmentation](https://docs.ultralytics.com/tasks/segment) seamlessly. ### YOLOv7: Pushing the Limits of Real-Time Accuracy @@ -78,7 +78,7 @@ While YOLOv7 achieves higher absolute mAP scores on larger variants, YOLOv5 offe A model's utility extends beyond its raw architecture; the ecosystem surrounding it dictates how quickly it can be deployed to production. This is where Ultralytics models shine. - **Ease of Use:** The [Ultralytics Platform](https://platform.ultralytics.com) and its unified Python API provide a streamlined user experience, simple syntax, and extensive documentation. Training a custom dataset requires zero boilerplate code. -- **Well-Maintained Ecosystem:** Ultralytics benefits from active development, frequent updates, and strong community support. Integrations with tools like [Comet ML](https://docs.ultralytics.com/integrations/comet/) and [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) are baked right in. +- **Well-Maintained Ecosystem:** Ultralytics benefits from active development, frequent updates, and strong community support. Integrations with tools like [Comet ML](https://docs.ultralytics.com/integrations/comet) and [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) are baked right in. - **Training Efficiency:** Data loaders, smart caching, and multi-GPU support make Ultralytics models exceptionally efficient to train. Readily available pre-trained weights dramatically accelerate [transfer learning](https://www.ultralytics.com/glossary/transfer-learning). ### Code Example: Getting Started @@ -114,7 +114,7 @@ YOLOv7 remains a strong candidate for academic benchmarking or specific legacy G YOLOv5 is heavily favored for production environments due to its exceptional stability. It is the go-to choice for: -- **Mobile and Edge Computing:** Deploying YOLOv5n to iOS via [CoreML](https://docs.ultralytics.com/integrations/coreml/) or Android via [TFLite](https://docs.ultralytics.com/integrations/tflite/). +- **Mobile and Edge Computing:** Deploying YOLOv5n to iOS via [CoreML](https://docs.ultralytics.com/integrations/coreml) or Android via [TFLite](https://docs.ultralytics.com/integrations/tflite). - **Agile Startups:** Teams needing rapid iteration cycles benefit from the seamless [Ultralytics Platform](https://platform.ultralytics.com) integration for dataset management and cloud training. - **Multi-Task Environments:** Systems requiring simultaneous object detection, classification, and segmentation. diff --git a/docs/en/compare/yolov5-vs-yolov8.md b/docs/en/compare/yolov5-vs-yolov8.md index e5cdaa392cc..c07ffe4a7db 100644 --- a/docs/en/compare/yolov5-vs-yolov8.md +++ b/docs/en/compare/yolov5-vs-yolov8.md @@ -13,7 +13,7 @@ When building scalable and efficient [computer vision](https://en.wikipedia.org/ -Both of these models represent significant milestones in the history of real-time [object detection](https://docs.ultralytics.com/tasks/detect/), and both benefit from the highly optimized memory requirements and [ease of use](https://docs.ultralytics.com/quickstart/) that characterize the Ultralytics ecosystem. +Both of these models represent significant milestones in the history of real-time [object detection](https://docs.ultralytics.com/tasks/detect), and both benefit from the highly optimized memory requirements and [ease of use](https://docs.ultralytics.com/quickstart) that characterize the Ultralytics ecosystem. ## YOLOv5: The Reliable Industry Standard @@ -23,7 +23,7 @@ Introduced in 2020, YOLOv5 rapidly became the industry standard for fast, access - **Organization:** [Ultralytics](https://www.ultralytics.com) - **Date:** 2020-06-26 - **GitHub:** [ultralytics/yolov5](https://github.com/ultralytics/yolov5) -- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5/) +- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5) ### Architectural Strengths @@ -31,7 +31,7 @@ YOLOv5 operates on an anchor-based detection paradigm, which relies on predefine ### Ideal Use Cases -YOLOv5 is highly recommended for projects where maximum throughput and minimal resource utilization are paramount. It excels in [edge AI](https://www.ultralytics.com/glossary/edge-ai) environments, such as deploying on [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or mobile devices. Its maturity means it has been thoroughly battle-tested in thousands of commercial deployments, offering unmatched stability for traditional object detection workflows. +YOLOv5 is highly recommended for projects where maximum throughput and minimal resource utilization are paramount. It excels in [edge AI](https://www.ultralytics.com/glossary/edge-ai) environments, such as deploying on [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or mobile devices. Its maturity means it has been thoroughly battle-tested in thousands of commercial deployments, offering unmatched stability for traditional object detection workflows. !!! tip "Legacy Deployment Advantage" @@ -47,17 +47,17 @@ Released in January 2023, YOLOv8 represented a monumental architectural shift, e - **Organization:** [Ultralytics](https://www.ultralytics.com) - **Date:** 2023-01-10 - **GitHub:** [ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8/) +- **Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8) ### Architectural Innovations Unlike its predecessor, YOLOv8 introduces an **anchor-free** detection head. This eliminates the need to manually tune anchor configurations based on dataset distributions, enhancing generalization across diverse custom datasets like the popular [COCO dataset](https://cocodataset.org/). -The architecture also upgrades the backbone with a **C2f module** (Cross-Stage Partial bottleneck with two convolutions), replacing the older C3 module. This enhancement improves feature representation without heavily taxing memory. Additionally, the implementation of a decoupled head—separating objectness, classification, and regression tasks—drastically improves convergence during [model training](https://docs.ultralytics.com/modes/train/). +The architecture also upgrades the backbone with a **C2f module** (Cross-Stage Partial bottleneck with two convolutions), replacing the older C3 module. This enhancement improves feature representation without heavily taxing memory. Additionally, the implementation of a decoupled head—separating objectness, classification, and regression tasks—drastically improves convergence during [model training](https://docs.ultralytics.com/modes/train). ### Versatility and Python API -YOLOv8 introduced the modern `ultralytics` Python API, standardizing the workflow across various computer vision tasks. Whether you are performing [image segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), or [pose estimation](https://docs.ultralytics.com/tasks/pose/), the unified API requires only minor configuration changes. +YOLOv8 introduced the modern `ultralytics` Python API, standardizing the workflow across various computer vision tasks. Whether you are performing [image segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), or [pose estimation](https://docs.ultralytics.com/tasks/pose), the unified API requires only minor configuration changes. ```python from ultralytics import YOLO @@ -77,7 +77,7 @@ predictions[0].show() ## Detailed Performance Comparison -When comparing the two generations, we observe a classic trade-off: YOLOv8 achieves higher mean Average Precision ([mAP](https://docs.ultralytics.com/guides/yolo-performance-metrics/)) across the board, while YOLOv5 retains a slight edge in absolute raw inference speed and parameter count for its smallest variants. +When comparing the two generations, we observe a classic trade-off: YOLOv8 achieves higher mean Average Precision ([mAP](https://docs.ultralytics.com/guides/yolo-performance-metrics)) across the board, while YOLOv5 retains a slight edge in absolute raw inference speed and parameter count for its smallest variants. Below is the detailed comparison of their performance metrics on the COCO dataset at an image size of 640 pixels. @@ -103,15 +103,15 @@ The data reveals that YOLOv8 provides a substantial boost in accuracy. For insta ## The Ecosystem Advantage -Choosing either YOLOv5 or YOLOv8 grants developers access to the well-maintained [Ultralytics Platform](https://platform.ultralytics.com/). This integrated environment offers simple tools for dataset annotation, [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/), cloud training, and model monitoring. The active development and strong community support ensure that developers can quickly resolve issues and integrate with external tools like [Weights & Biases](https://wandb.ai/site) and [ClearML](https://clear.ml/). +Choosing either YOLOv5 or YOLOv8 grants developers access to the well-maintained [Ultralytics Platform](https://platform.ultralytics.com/). This integrated environment offers simple tools for dataset annotation, [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning), cloud training, and model monitoring. The active development and strong community support ensure that developers can quickly resolve issues and integrate with external tools like [Weights & Biases](https://wandb.ai/site) and [ClearML](https://clear.ml/). While other frameworks might suffer from steep learning curves, Ultralytics prioritizes a streamlined user experience, ensuring a favorable trade-off between speed and accuracy suitable for diverse real-world deployment scenarios. ## Beyond v8: Exploring YOLO11 and YOLO26 -While YOLOv8 is a highly capable framework, the field of artificial intelligence evolves rapidly. Developers interested in state-of-the-art performance should also explore [YOLO11](https://docs.ultralytics.com/models/yolo11/), which builds upon v8 with improved precision and speed. +While YOLOv8 is a highly capable framework, the field of artificial intelligence evolves rapidly. Developers interested in state-of-the-art performance should also explore [YOLO11](https://docs.ultralytics.com/models/yolo11), which builds upon v8 with improved precision and speed. -For those seeking the absolute bleeding edge of computer vision technology, we highly recommend **[Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/)**. Released in 2026, YOLO26 represents a massive leap forward: +For those seeking the absolute bleeding edge of computer vision technology, we highly recommend **[Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26)**. Released in 2026, YOLO26 represents a massive leap forward: - **End-to-End NMS-Free Design:** Pioneered originally in experimental architectures, YOLO26 natively eliminates Non-Maximum Suppression post-processing, leading to drastically simpler and faster deployment pipelines. - **MuSGD Optimizer:** Inspired by the LLM training innovations seen in models like Kimi K2, YOLO26 utilizes a hybrid optimizer for more stable training and rapid convergence. diff --git a/docs/en/compare/yolov5-vs-yolov9.md b/docs/en/compare/yolov5-vs-yolov9.md index dd61520cc19..3defde56762 100644 --- a/docs/en/compare/yolov5-vs-yolov9.md +++ b/docs/en/compare/yolov5-vs-yolov9.md @@ -8,7 +8,7 @@ keywords: YOLOv5, YOLOv9, object detection, model comparison, performance metric The landscape of computer vision and real-time object detection has seen remarkable advancements over the past few years. Navigating the choice between established, battle-tested models and newer research architectures is a common challenge for machine learning engineers. This guide provides a comprehensive technical comparison between two highly influential models in the YOLO family: **YOLOv5** and **YOLOv9**. -Whether you are deploying to constrained edge devices, researching high-fidelity feature extraction, or building complex [object detection](https://docs.ultralytics.com/tasks/detect/) pipelines, understanding the architectural nuances, performance metrics, and ecosystem differences of these models is crucial. +Whether you are deploying to constrained edge devices, researching high-fidelity feature extraction, or building complex [object detection](https://docs.ultralytics.com/tasks/detect) pipelines, understanding the architectural nuances, performance metrics, and ecosystem differences of these models is crucial. @@ -29,7 +29,7 @@ Developed by Glenn Jocher and released by [Ultralytics](https://www.ultralytics. - **GitHub:** [YOLOv5 Repository](https://github.com/ultralytics/yolov5) - **Docs:** [YOLOv5 Platform Overview](https://platform.ultralytics.com/ultralytics/yolov5) -YOLOv5 is renowned for its **Ease of Use** and stable performance across diverse hardware environments. It supports not just detection, but also [image classification](https://docs.ultralytics.com/tasks/classify/) and [instance segmentation](https://docs.ultralytics.com/tasks/segment/). +YOLOv5 is renowned for its **Ease of Use** and stable performance across diverse hardware environments. It supports not just detection, but also [image classification](https://docs.ultralytics.com/tasks/classify) and [instance segmentation](https://docs.ultralytics.com/tasks/segment). [Learn more about YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5){ .md-button } @@ -42,11 +42,11 @@ Introduced by Chien-Yao Wang and Hong-Yuan Mark Liao from the Institute of Infor - **Date:** 2024-02-21 - **Arxiv:** [2402.13616](https://arxiv.org/abs/2402.13616) - **GitHub:** [YOLOv9 Repository](https://github.com/WongKinYiu/yolov9) -- **Docs:** [YOLOv9 Documentation](https://docs.ultralytics.com/models/yolov9/) +- **Docs:** [YOLOv9 Documentation](https://docs.ultralytics.com/models/yolov9) The core of YOLOv9 relies on two major theoretical innovations: Programmable Gradient Information (PGI) and the Generalized Efficient Layer Aggregation Network (GELAN). These concepts help the model retain critical spatial features through deep network layers. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } !!! tip "Future-Proof Your Deployments" @@ -54,7 +54,7 @@ The core of YOLOv9 relies on two major theoretical innovations: Programmable Gra ## Architectural and Technical Differences -Understanding what powers these vision models under the hood is vital for optimizing [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/) strategies. +Understanding what powers these vision models under the hood is vital for optimizing [model deployment](https://docs.ultralytics.com/guides/model-deployment-options) strategies. ### Feature Extraction and Information Retention @@ -92,7 +92,7 @@ The true advantage of leveraging [computer vision](https://www.ultralytics.com/g ### The Ultralytics Advantage -While original research repositories for models like YOLOv9 are foundational, they often come with complex dependency matrices and boilerplate scripts. The [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) completely abstracts this complexity. With the Ultralytics ecosystem, you can train, evaluate, and export both YOLOv5 and YOLOv9 with an identical, unified syntax. +While original research repositories for models like YOLOv9 are foundational, they often come with complex dependency matrices and boilerplate scripts. The [Ultralytics Python API](https://docs.ultralytics.com/usage/python) completely abstracts this complexity. With the Ultralytics ecosystem, you can train, evaluate, and export both YOLOv5 and YOLOv9 with an identical, unified syntax. ```python from ultralytics import YOLO @@ -110,7 +110,7 @@ results = model_v9.train(data="coco8.yaml", epochs=50, imgsz=640) model_v9.export(format="onnx") ``` -This single-API approach provides immense **Versatility**, supporting not just detection, but [pose estimation](https://docs.ultralytics.com/tasks/pose/) and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) depending on the model chosen. Furthermore, robust integrations with tools like [Comet ML](https://docs.ultralytics.com/integrations/comet/) and [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) are baked directly into the training loop. +This single-API approach provides immense **Versatility**, supporting not just detection, but [pose estimation](https://docs.ultralytics.com/tasks/pose) and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb) depending on the model chosen. Furthermore, robust integrations with tools like [Comet ML](https://docs.ultralytics.com/integrations/comet) and [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) are baked directly into the training loop. ## Ideal Use Cases and Real-World Applications @@ -120,8 +120,8 @@ Choosing between these architectures depends largely on the constraints of your YOLOv5 is a battle-hardened veteran that shines in deployments prioritizing stability, low memory footprints, and extreme export compatibility. -- **Mobile Deployments:** Exporting YOLOv5 to [TFLite](https://docs.ultralytics.com/integrations/tflite/) or CoreML for on-device inference on older smartphones is incredibly seamless. -- **Legacy Edge Hardware:** For devices like the Raspberry Pi or early generation NVIDIA Jetson Nanos, the straightforward convolutions of YOLOv5 ensure consistent frame rates for applications like [smart parking management](https://docs.ultralytics.com/guides/parking-management/). +- **Mobile Deployments:** Exporting YOLOv5 to [TFLite](https://docs.ultralytics.com/integrations/tflite) or CoreML for on-device inference on older smartphones is incredibly seamless. +- **Legacy Edge Hardware:** For devices like the Raspberry Pi or early generation NVIDIA Jetson Nanos, the straightforward convolutions of YOLOv5 ensure consistent frame rates for applications like [smart parking management](https://docs.ultralytics.com/guides/parking-management). - **Rapid Prototyping:** The extensive availability of community tutorials, custom [pre-trained weights](https://www.ultralytics.com/glossary/model-weights), and massive dataset compatibility makes it the fastest way to validate a proof-of-concept. ### When to Choose YOLOv9 @@ -134,6 +134,6 @@ YOLOv9 is ideal for scenarios where capturing intricate details and minimizing f ## Expanding Your Horizons -While comparing YOLOv5 and YOLOv9 offers a clear view of how architectures have evolved from 2020 to 2024, the field of AI is moving faster than ever. For developers seeking the absolute frontier of performance, exploring the latest [YOLO26 models](https://docs.ultralytics.com/models/yolo26/) is highly encouraged. By replacing traditional Non-Maximum Suppression with a native **End-to-End NMS-Free Design** and utilizing the advanced **MuSGD Optimizer**, YOLO26 bridges the gap between research-level accuracy and production-level speed. With **DFL Removal** (Distribution Focal Loss removed for simplified export and better edge/low-power device compatibility), YOLO26 achieves up to **43% faster CPU inference**, making it ideal for edge computing. Additionally, **ProgLoss + STAL** provides improved loss functions with notable improvements in small-object recognition, critical for IoT, robotics, and aerial imagery. +While comparing YOLOv5 and YOLOv9 offers a clear view of how architectures have evolved from 2020 to 2024, the field of AI is moving faster than ever. For developers seeking the absolute frontier of performance, exploring the latest [YOLO26 models](https://docs.ultralytics.com/models/yolo26) is highly encouraged. By replacing traditional Non-Maximum Suppression with a native **End-to-End NMS-Free Design** and utilizing the advanced **MuSGD Optimizer**, YOLO26 bridges the gap between research-level accuracy and production-level speed. With **DFL Removal** (Distribution Focal Loss removed for simplified export and better edge/low-power device compatibility), YOLO26 achieves up to **43% faster CPU inference**, making it ideal for edge computing. Additionally, **ProgLoss + STAL** provides improved loss functions with notable improvements in small-object recognition, critical for IoT, robotics, and aerial imagery. -You might also be interested in comparing these architectures against other state-of-the-art models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) or the highly capable [YOLO11](https://docs.ultralytics.com/models/yolo11/). Utilizing the unified Ultralytics framework ensures that no matter which model you choose, your development pipeline remains clean, efficient, and ready to scale. +You might also be interested in comparing these architectures against other state-of-the-art models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) or the highly capable [YOLO11](https://docs.ultralytics.com/models/yolo11). Utilizing the unified Ultralytics framework ensures that no matter which model you choose, your development pipeline remains clean, efficient, and ready to scale. diff --git a/docs/en/compare/yolov5-vs-yolox.md b/docs/en/compare/yolov5-vs-yolox.md index f4b60177bda..dc386d4f0db 100644 --- a/docs/en/compare/yolov5-vs-yolox.md +++ b/docs/en/compare/yolov5-vs-yolox.md @@ -23,7 +23,7 @@ This guide provides an in-depth technical analysis of these two models, comparin - **Organization:** [Ultralytics](https://www.ultralytics.com) - **Date:** 2020-06-26 - **GitHub:** [Ultralytics YOLOv5 Repository](https://github.com/ultralytics/yolov5) -- **Documentation:** [YOLOv5 Official Docs](https://docs.ultralytics.com/models/yolov5/) +- **Documentation:** [YOLOv5 Official Docs](https://docs.ultralytics.com/models/yolov5) Introduced by Ultralytics, [YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5) quickly became an industry standard due to its exceptional balance of performance, ease of use, and memory efficiency. Built natively on the [PyTorch](https://pytorch.org/) framework, YOLOv5 uses an anchor-based architecture. It relies on predefined bounding box shapes to predict object locations, which makes it highly effective for standard object detection tasks. @@ -96,7 +96,7 @@ results = model.train(data="coco8.yaml", epochs=100, imgsz=640) metrics = model.val() ``` -Furthermore, YOLOv5's versatility extends beyond standard object detection, offering robust support for [image classification](https://docs.ultralytics.com/tasks/classify/) and [instance segmentation](https://docs.ultralytics.com/tasks/segment/) within the exact same cohesive API. +Furthermore, YOLOv5's versatility extends beyond standard object detection, offering robust support for [image classification](https://docs.ultralytics.com/tasks/classify) and [instance segmentation](https://docs.ultralytics.com/tasks/segment) within the exact same cohesive API. !!! tip "Streamlined Deployment" @@ -108,7 +108,7 @@ Choosing between these models depends on your deployment environment and technic - **Retail and Inventory Management:** For applications requiring real-time product recognition on edge devices like the NVIDIA Jetson, **YOLOv5** is exceptionally well-suited. Its minimal memory footprint and fast TensorRT inference speeds enable multi-camera tracking without dropping frames. - **Academic Research and Custom Architectures:** **YOLOX** is highly regarded in the research community. Its decoupled head and anchor-free nature make it an excellent baseline for engineers looking to experiment with novel label assignment strategies or those working on datasets where traditional anchor boxes fail to generalize. -- **Agricultural AI:** For precision agriculture tasks like fruit detection or weed identification via drones, the ease of training and deploying YOLOv5 models using the [Ultralytics Platform](https://docs.ultralytics.com/platform/) allows domain experts to implement AI solutions without needing deep machine learning engineering backgrounds. +- **Agricultural AI:** For precision agriculture tasks like fruit detection or weed identification via drones, the ease of training and deploying YOLOv5 models using the [Ultralytics Platform](https://docs.ultralytics.com/platform) allows domain experts to implement AI solutions without needing deep machine learning engineering backgrounds. ## Use Cases and Recommendations @@ -120,7 +120,7 @@ YOLOv5 is a strong choice for: - **Proven Production Systems:** Existing deployments where YOLOv5's long track record of stability, extensive documentation, and massive community support are valued. - **Resource-Constrained Training:** Environments with limited GPU resources where YOLOv5's efficient training pipeline and lower memory requirements are advantageous. -- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), [CoreML](https://docs.ultralytics.com/integrations/coreml/), and [TFLite](https://docs.ultralytics.com/integrations/tflite/). +- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), [CoreML](https://docs.ultralytics.com/integrations/coreml), and [TFLite](https://docs.ultralytics.com/integrations/tflite). ### When to Choose YOLOX @@ -132,11 +132,11 @@ YOLOX is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Future of Vision AI: Enter YOLO26 diff --git a/docs/en/compare/yolov6-vs-damo-yolo.md b/docs/en/compare/yolov6-vs-damo-yolo.md index d94f5f2d746..e38463112be 100644 --- a/docs/en/compare/yolov6-vs-damo-yolo.md +++ b/docs/en/compare/yolov6-vs-damo-yolo.md @@ -6,7 +6,7 @@ keywords: YOLOv6-3.0, DAMO-YOLO, object detection comparison, YOLO models, compu # YOLOv6-3.0 vs DAMO-YOLO: A Technical Showdown in Real-Time Object Detection -The landscape of computer vision is constantly evolving, with new architectures pushing the boundaries of what is possible in real-time [object detection](https://docs.ultralytics.com/tasks/detect/). Two notable contenders in this space are YOLOv6-3.0 and DAMO-YOLO. Both models introduce unique architectural innovations designed to maximize performance on industrial hardware. This guide provides a comprehensive technical comparison between these two models, exploring their architectures, training methodologies, and ideal use cases, while also introducing the next-generation advantages of Ultralytics models like YOLO26. +The landscape of computer vision is constantly evolving, with new architectures pushing the boundaries of what is possible in real-time [object detection](https://docs.ultralytics.com/tasks/detect). Two notable contenders in this space are YOLOv6-3.0 and DAMO-YOLO. Both models introduce unique architectural innovations designed to maximize performance on industrial hardware. This guide provides a comprehensive technical comparison between these two models, exploring their architectures, training methodologies, and ideal use cases, while also introducing the next-generation advantages of Ultralytics models like YOLO26. @@ -24,11 +24,11 @@ Developed by the Vision AI Department at [Meituan](https://tech.meituan.com/), Y - **Date:** 2023-01-13 - **Arxiv:** [2301.05586](https://arxiv.org/abs/2301.05586) - **GitHub:** [meituan/YOLOv6](https://github.com/meituan/YOLOv6) -- **Docs:** [Ultralytics YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6/) +- **Docs:** [Ultralytics YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6) YOLOv6-3.0 introduces a Bi-directional Concatenation (BiC) module to improve feature fusion and utilizes an Anchor-Aided Training (AAT) strategy. This strategy combines the benefits of anchor-based and [anchor-free detectors](https://www.ultralytics.com/glossary/anchor-free-detectors) during training, while keeping inference strictly anchor-free. Its EfficientRep backbone makes it highly hardware-friendly for GPU batch processing, ideal for processing vast amounts of [video understanding](https://www.ultralytics.com/glossary/video-understanding) data. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ### DAMO-YOLO: Fast and Accurate via NAS @@ -74,18 +74,18 @@ With the release of **YOLO26**, Ultralytics has redefined state-of-the-art visio ### Key Innovations in YOLO26 -- **End-to-End NMS-Free Design:** Building on concepts pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing. This drastically reduces latency variance and simplifies deployment on edge devices via [CoreML](https://developer.apple.com/machine-learning/core-ml/) or [TFLite](https://ai.google.dev/edge/litert). +- **End-to-End NMS-Free Design:** Building on concepts pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing. This drastically reduces latency variance and simplifies deployment on edge devices via [CoreML](https://developer.apple.com/machine-learning/core-ml/) or [TFLite](https://ai.google.dev/edge/litert). - **DFL Removal:** By removing Distribution Focal Loss, YOLO26 simplifies the export process and significantly enhances compatibility with low-power microcontrollers and edge hardware. - **Up to 43% Faster CPU Inference:** For applications lacking dedicated GPU hardware, YOLO26's CPU optimizations deliver unparalleled speed, outperforming heavily GPU-reliant models like YOLOv6. - **MuSGD Optimizer:** Inspired by LLM training techniques like Moonshot AI's Kimi K2, YOLO26 utilizes the MuSGD optimizer (a hybrid of SGD and Muon) to guarantee stable training and rapid convergence. - **ProgLoss + STAL:** Advanced loss functions dramatically improve small-object recognition, making YOLO26 perfect for [drone operations](https://www.ultralytics.com/blog/computer-vision-applications-ai-drone-uav-operations) and distant target tracking. -- **Multi-Task Versatility:** Unlike DAMO-YOLO, which is strictly a detector, YOLO26 provides out-of-the-box support for [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/) (via Residual Log-Likelihood Estimation), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) within a single, unified API. +- **Multi-Task Versatility:** Unlike DAMO-YOLO, which is strictly a detector, YOLO26 provides out-of-the-box support for [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose) (via Residual Log-Likelihood Estimation), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb) within a single, unified API. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } !!! tip "Memory Efficient Training" - Unlike complex transformer architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) or the distillation-heavy pipelines of DAMO-YOLO, Ultralytics models are renowned for their low VRAM footprint. You can easily train a YOLO26 model on consumer-grade hardware. + Unlike complex transformer architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) or the distillation-heavy pipelines of DAMO-YOLO, Ultralytics models are renowned for their low VRAM footprint. You can easily train a YOLO26 model on consumer-grade hardware. ### Streamlined Python Workflow @@ -124,7 +124,7 @@ Choosing the right architecture depends entirely on your deployment constraints: ### When to use Ultralytics YOLO26 -- **Edge and Mobile Deployments:** The NMS-free design, DFL removal, and 43% CPU speed boost make it the undisputed champion for iOS, Android, and [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) integrations. +- **Edge and Mobile Deployments:** The NMS-free design, DFL removal, and 43% CPU speed boost make it the undisputed champion for iOS, Android, and [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) integrations. - **Rapid Prototyping to Production:** The seamless integration with the [Ultralytics Platform](https://platform.ultralytics.com/) enables teams to go from dataset annotation to global cloud deployment in days, not months. - **Complex Vision Pipelines:** When a project requires detecting bounding boxes alongside human pose keypoints and precise segmentation masks simultaneously. @@ -134,4 +134,4 @@ Both YOLOv6-3.0 and DAMO-YOLO have contributed significantly to the science of r However, for developers seeking the ultimate blend of accuracy, inference speed, and ecosystem maintainability, the [Ultralytics YOLO](https://www.ultralytics.com/) family remains the premier choice. With the groundbreaking optimizations introduced in **YOLO26**, the barrier to entry for creating enterprise-grade computer vision applications has never been lower. -For further exploration, you might also be interested in comparing these models to other architectures in our documentation, such as [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or transformer-based approaches like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +For further exploration, you might also be interested in comparing these models to other architectures in our documentation, such as [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or transformer-based approaches like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). diff --git a/docs/en/compare/yolov6-vs-efficientdet.md b/docs/en/compare/yolov6-vs-efficientdet.md index 1a10c7cbbd7..810a3d35a30 100644 --- a/docs/en/compare/yolov6-vs-efficientdet.md +++ b/docs/en/compare/yolov6-vs-efficientdet.md @@ -6,7 +6,7 @@ keywords: YOLOv6, EfficientDet, object detection, model comparison, YOLOv6-3.0, # YOLOv6-3.0 vs. EfficientDet: A Comprehensive Technical Comparison -Choosing the optimal architecture for [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) projects requires a deep understanding of the trade-offs between speed, accuracy, and deployment feasibility. This comparison page provides an in-depth analysis of two distinct object detection models: YOLOv6-3.0 and EfficientDet. While both models have contributed significantly to the field, modern edge deployments and rapid prototyping often benefit from more unified frameworks like the [Ultralytics Platform](https://docs.ultralytics.com/platform/). +Choosing the optimal architecture for [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) projects requires a deep understanding of the trade-offs between speed, accuracy, and deployment feasibility. This comparison page provides an in-depth analysis of two distinct object detection models: YOLOv6-3.0 and EfficientDet. While both models have contributed significantly to the field, modern edge deployments and rapid prototyping often benefit from more unified frameworks like the [Ultralytics Platform](https://docs.ultralytics.com/platform). Below is an interactive chart visualizing the performance differences between these models to help you understand their respective latency and accuracy profiles. @@ -33,7 +33,7 @@ The YOLOv6-3.0 architecture relies on a Bi-directional Concatenation (BiC) modul YOLOv6-3.0 shines in environments where dedicated GPU hardware is available, offering incredibly fast [real-time inference](https://www.ultralytics.com/glossary/real-time-inference) using TensorRT. However, its heavy reliance on specific hardware optimizations can lead to suboptimal performance on CPU-only [edge AI devices](https://www.ultralytics.com/glossary/edge-ai). Additionally, while it supports some quantization, the ecosystem lacks the overarching simplicity found in modern Ultralytics frameworks. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## EfficientDet: Scalable AutoML Architecture @@ -51,7 +51,7 @@ EfficientDet introduced the Bi-directional Feature Pyramid Network (BiFPN), whic ### Strengths and Weaknesses -EfficientDet is highly parameter-efficient. It achieves strong [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map) with relatively few parameters compared to older object detectors. However, the architecture is deeply entrenched in legacy TensorFlow ecosystems. This results in complex dependency management, slower training cycles, and higher [memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics/) during training compared to optimized PyTorch implementations. Furthermore, its inference speed on modern GPUs is significantly slower than modern YOLO architectures. +EfficientDet is highly parameter-efficient. It achieves strong [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map) with relatively few parameters compared to older object detectors. However, the architecture is deeply entrenched in legacy TensorFlow ecosystems. This results in complex dependency management, slower training cycles, and higher [memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics) during training compared to optimized PyTorch implementations. Furthermore, its inference speed on modern GPUs is significantly slower than modern YOLO architectures. [Learn more about EfficientDet](https://github.com/google/automl/tree/master/efficientdet){ .md-button } @@ -89,7 +89,7 @@ Unlike EfficientDet, which requires navigating complex TensorFlow configurations ### Unmatched Versatility -YOLOv6-3.0 and EfficientDet are primarily bound to [object detection](https://docs.ultralytics.com/tasks/detect/). In contrast, modern Ultralytics architectures are inherently multi-modal. A single interface allows you to train models for [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) tasks. +YOLOv6-3.0 and EfficientDet are primarily bound to [object detection](https://docs.ultralytics.com/tasks/detect). In contrast, modern Ultralytics architectures are inherently multi-modal. A single interface allows you to train models for [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Image Classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) tasks. ### Introducing Ultralytics YOLO26 @@ -124,11 +124,11 @@ EfficientDet is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Implementation Example: Training YOLO26 @@ -156,11 +156,11 @@ prediction = model("https://ultralytics.com/images/bus.jpg") If you are exploring the broader landscape of computer vision models, consider these alternatives: - **[YOLO11](https://platform.ultralytics.com/ultralytics/yolo11):** The highly successful predecessor to YOLO26, offering robust multi-task capabilities and extensive community support. -- **[YOLOv10](https://docs.ultralytics.com/models/yolov10/):** The first YOLO architecture to introduce NMS-free training, paving the way for modern end-to-end detection. -- **[RT-DETR](https://docs.ultralytics.com/models/rtdetr/):** For scenarios where transformer-based architectures and attention mechanisms are preferred over traditional CNNs. +- **[YOLOv10](https://docs.ultralytics.com/models/yolov10):** The first YOLO architecture to introduce NMS-free training, paving the way for modern end-to-end detection. +- **[RT-DETR](https://docs.ultralytics.com/models/rtdetr):** For scenarios where transformer-based architectures and attention mechanisms are preferred over traditional CNNs. ## Conclusion While **YOLOv6-3.0** provides excellent industrial GPU throughput and **EfficientDet** showcases the potential of AutoML in crafting scalable parameter-efficient networks, both models exhibit limitations in ease of deployment and modern multi-task versatility. -For the vast majority of real-world applications—from mobile edge deployment to cloud-based analytics—the **Ultralytics ecosystem** delivers an unparalleled [performance balance](https://docs.ultralytics.com/guides/yolo-performance-metrics/). By adopting **YOLO26**, developers gain access to cutting-edge NMS-free inference, advanced loss functions for small objects, and a unified, well-documented training pipeline that dramatically accelerates the path from prototype to production. +For the vast majority of real-world applications—from mobile edge deployment to cloud-based analytics—the **Ultralytics ecosystem** delivers an unparalleled [performance balance](https://docs.ultralytics.com/guides/yolo-performance-metrics). By adopting **YOLO26**, developers gain access to cutting-edge NMS-free inference, advanced loss functions for small objects, and a unified, well-documented training pipeline that dramatically accelerates the path from prototype to production. diff --git a/docs/en/compare/yolov6-vs-pp-yoloe.md b/docs/en/compare/yolov6-vs-pp-yoloe.md index a0cc60dd001..4594dd368e9 100644 --- a/docs/en/compare/yolov6-vs-pp-yoloe.md +++ b/docs/en/compare/yolov6-vs-pp-yoloe.md @@ -6,7 +6,7 @@ keywords: YOLOv6-3.0, PP-YOLOE+, object detection, model comparison, computer vi # YOLOv6-3.0 vs PP-YOLOE+: Evaluating Industrial Object Detectors -When selecting a framework for real-time [object detection](https://docs.ultralytics.com/tasks/detect/), machine learning engineers frequently evaluate a variety of high-performance architectures. Two notable models in the landscape of industrial applications are **YOLOv6-3.0** and **PP-YOLOE+**. Both models have pushed the boundaries of accuracy and speed, yet they are tailored for slightly different ecosystems and deployment hardware. +When selecting a framework for real-time [object detection](https://docs.ultralytics.com/tasks/detect), machine learning engineers frequently evaluate a variety of high-performance architectures. Two notable models in the landscape of industrial applications are **YOLOv6-3.0** and **PP-YOLOE+**. Both models have pushed the boundaries of accuracy and speed, yet they are tailored for slightly different ecosystems and deployment hardware. @@ -29,7 +29,7 @@ Developed by the Vision AI Department at **Meituan**, YOLOv6-3.0 is heavily opti YOLOv6-3.0 utilizes an **EfficientRep** backbone, specifically designed to maximize utilization of hardware accelerators like NVIDIA GPUs. The architecture introduces a **Bi-directional Concatenation (BiC)** module within the neck, significantly improving the fusion of multi-scale features. Furthermore, it incorporates an **Anchor-Aided Training (AAT)** strategy. This hybrid approach enjoys the robust convergence characteristics of [anchor-based networks](https://www.ultralytics.com/glossary/anchor-boxes) during the training phase, while discarding the anchors during inference to maintain the high speed typical of anchor-free paradigms. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## PP-YOLOE+: PaddlePaddle's Detection Champion @@ -43,7 +43,7 @@ YOLOv6-3.0 utilizes an **EfficientRep** backbone, specifically designed to maxim ### Architectural Innovations -PP-YOLOE+ is an **anchor-free** detector that introduces a dynamic label assignment strategy known as TAL (Task Alignment Learning). It utilizes a CSPRepResNet backbone, which efficiently captures semantic features while maintaining computational efficiency. The model is highly optimized for deployment via TensorRT and OpenVINO, making it a strong contender for edge and server deployments, provided the user is comfortable navigating the [PaddlePaddle API](https://docs.ultralytics.com/integrations/paddlepaddle/). +PP-YOLOE+ is an **anchor-free** detector that introduces a dynamic label assignment strategy known as TAL (Task Alignment Learning). It utilizes a CSPRepResNet backbone, which efficiently captures semantic features while maintaining computational efficiency. The model is highly optimized for deployment via TensorRT and OpenVINO, making it a strong contender for edge and server deployments, provided the user is comfortable navigating the [PaddlePaddle API](https://docs.ultralytics.com/integrations/paddlepaddle). [Learn more about PP-YOLOE+](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.8.1/configs/ppyoloe/README.md){ .md-button } @@ -53,7 +53,7 @@ PP-YOLOE+ is an **anchor-free** detector that introduces a dynamic label assignm ## Performance Comparison -Evaluating these models requires looking at their balance of [mean average precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics/) and inference speed. The table below highlights their performance on the COCO validation dataset. +Evaluating these models requires looking at their balance of [mean average precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics) and inference speed. The table below highlights their performance on the COCO validation dataset. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ----------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -72,15 +72,15 @@ While both models show strong performance, YOLOv6-3.0 generally maintains a slig ## The Ultralytics Advantage: Introducing YOLO26 -While YOLOv6-3.0 and PP-YOLOE+ are highly capable, the rapid evolution of [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) demands architectures that offer not just raw speed, but also exceptional ease of use, lower memory requirements, and a unified ecosystem. This is where **Ultralytics YOLO** models, particularly [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and the cutting-edge **[YOLO26](https://docs.ultralytics.com/models/yolo26/)**, redefine the state-of-the-art. +While YOLOv6-3.0 and PP-YOLOE+ are highly capable, the rapid evolution of [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) demands architectures that offer not just raw speed, but also exceptional ease of use, lower memory requirements, and a unified ecosystem. This is where **Ultralytics YOLO** models, particularly [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and the cutting-edge **[YOLO26](https://docs.ultralytics.com/models/yolo26)**, redefine the state-of-the-art. Released in January 2026, **YOLO26** establishes a new benchmark for edge-first and cloud-ready vision AI, offering significant advantages over legacy models: -- **End-to-End NMS-Free Design:** Building on the foundations laid by [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates Non-Maximum Suppression (NMS) during post-processing. This significantly simplifies deployment logic and reduces latency variability in crowded scenes. +- **End-to-End NMS-Free Design:** Building on the foundations laid by [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates Non-Maximum Suppression (NMS) during post-processing. This significantly simplifies deployment logic and reduces latency variability in crowded scenes. - **Up to 43% Faster CPU Inference:** By strategically removing Distribution Focal Loss (DFL), YOLO26 drastically accelerates CPU performance, making it vastly superior to YOLOv6 or PP-YOLOE+ for IoT devices and mobile applications. - **MuSGD Optimizer:** Inspired by advanced LLM training techniques (like Moonshot AI's Kimi K2), the hybrid **MuSGD** optimizer delivers incredibly stable and efficient training, converging faster than traditional SGD or AdamW. -- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, a critical factor for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and aerial surveillance. -- **Versatility Across Tasks:** Unlike YOLOv6-3.0 which is heavily focused on detection, YOLO26 supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection out-of-the-box. +- **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, a critical factor for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and aerial surveillance. +- **Versatility Across Tasks:** Unlike YOLOv6-3.0 which is heavily focused on detection, YOLO26 supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection out-of-the-box. ### Streamlined Training Ecosystem @@ -104,7 +104,7 @@ metrics = model.val() path = model.export(format="engine") ``` -This simple API, combined with lower memory usage during training compared to transformer-heavy models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/), democratizes high-performance AI. +This simple API, combined with lower memory usage during training compared to transformer-heavy models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr), democratizes high-performance AI. ## Ideal Use Cases and Deployment Strategies @@ -113,7 +113,7 @@ Choosing the right model dictates the success of your deployment pipeline. ### When to use YOLOv6-3.0 - **High-Speed Manufacturing:** Environments where industrial cameras feed directly into dedicated NVIDIA T4 or A100 GPUs, requiring consistent inference under 5ms. -- **Server-Side Video Analytics:** Processing multiple dense video streams where pure [GPU throughput](https://docs.ultralytics.com/guides/optimizing-openvino-latency-vs-throughput-modes/) is the primary bottleneck. +- **Server-Side Video Analytics:** Processing multiple dense video streams where pure [GPU throughput](https://docs.ultralytics.com/guides/optimizing-openvino-latency-vs-throughput-modes) is the primary bottleneck. ### When to use PP-YOLOE+ @@ -123,7 +123,7 @@ Choosing the right model dictates the success of your deployment pipeline. ### When to Choose Ultralytics YOLO26 - **Edge and IoT Devices:** With its NMS-free design and DFL removal, YOLO26 is the undisputed choice for deployments on Raspberry Pi, NXP, or mobile CPUs. -- **Multi-Task Applications:** Projects requiring simultaneous [object tracking](https://docs.ultralytics.com/modes/track/), pose estimation, or segmentation using a unified API. -- **Rapid Prototyping to Production:** Teams leveraging the [Ultralytics Platform](https://platform.ultralytics.com/) for streamlined [dataset annotation](https://docs.ultralytics.com/platform/data/annotation/), hyperparameter tuning, and one-click [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/). +- **Multi-Task Applications:** Projects requiring simultaneous [object tracking](https://docs.ultralytics.com/modes/track), pose estimation, or segmentation using a unified API. +- **Rapid Prototyping to Production:** Teams leveraging the [Ultralytics Platform](https://platform.ultralytics.com/) for streamlined [dataset annotation](https://docs.ultralytics.com/platform/data/annotation), hyperparameter tuning, and one-click [model deployment](https://docs.ultralytics.com/guides/model-deployment-options). -For developers looking to explore the broader landscape of detection models, frameworks like [YOLOX](https://github.com/Megvii-BaseDetection/YOLOX) and [DAMO-YOLO](https://docs.ultralytics.com/compare/damo-yolo-vs-yolov6/) also offer unique architectural approaches worth reviewing in the Ultralytics documentation. +For developers looking to explore the broader landscape of detection models, frameworks like [YOLOX](https://github.com/Megvii-BaseDetection/YOLOX) and [DAMO-YOLO](https://docs.ultralytics.com/compare/damo-yolo-vs-yolov6) also offer unique architectural approaches worth reviewing in the Ultralytics documentation. diff --git a/docs/en/compare/yolov6-vs-rtdetr.md b/docs/en/compare/yolov6-vs-rtdetr.md index 2d5ae3cd7da..a097bdfed89 100644 --- a/docs/en/compare/yolov6-vs-rtdetr.md +++ b/docs/en/compare/yolov6-vs-rtdetr.md @@ -45,7 +45,7 @@ YOLOv6-3.0 adopts a hardware-friendly **EfficientRep** backbone specifically tai - Heavier reliance on complex Non-Maximum Suppression (NMS) during post-processing, increasing latency variance. - Less active ecosystem compared to mainstream frameworks, making updates and community support less predictable. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } --- @@ -61,7 +61,7 @@ Spearheaded by researchers at Baidu, RTDETRv2 builds upon the original RT-DETR b ### Architecture Highlights -Unlike traditional CNNs, RTDETRv2 is natively end-to-end. By leveraging transformer attention layers, the architecture completely eliminates the need for NMS post-processing. This allows for a streamlined inference pipeline. RTDETRv2 introduces highly optimized cross-scale feature fusion and an efficient hybrid encoder, allowing it to process standard [COCO datasets](https://docs.ultralytics.com/datasets/detect/coco/) with remarkable precision. +Unlike traditional CNNs, RTDETRv2 is natively end-to-end. By leveraging transformer attention layers, the architecture completely eliminates the need for NMS post-processing. This allows for a streamlined inference pipeline. RTDETRv2 introduces highly optimized cross-scale feature fusion and an efficient hybrid encoder, allowing it to process standard [COCO datasets](https://docs.ultralytics.com/datasets/detect/coco) with remarkable precision. ### Strengths and Weaknesses @@ -77,7 +77,7 @@ Unlike traditional CNNs, RTDETRv2 is natively end-to-end. By leveraging transfor - CPU inference speeds are notably slower than specialized edge CNNs, limiting its use in mobile or IoT devices. - Setup and tuning can be complex for teams accustomed to traditional [machine learning operations (MLOps)](https://www.ultralytics.com/glossary/machine-learning-operations-mlops). -[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr){ .md-button } --- @@ -99,7 +99,7 @@ The following table benchmarks YOLOv6-3.0 and RTDETRv2 across key performance in !!! tip "Deployment Tip" - If you are deploying on strictly CPU hardware like a Raspberry Pi, CNN-based models generally far outperform transformer architectures in Frames Per Second (FPS). For optimal edge performance, consider utilizing [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) to accelerate your inference. + If you are deploying on strictly CPU hardware like a Raspberry Pi, CNN-based models generally far outperform transformer architectures in Frames Per Second (FPS). For optimal edge performance, consider utilizing [OpenVINO](https://docs.ultralytics.com/integrations/openvino) to accelerate your inference. --- @@ -125,11 +125,11 @@ RT-DETR is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Enter YOLO26 @@ -139,11 +139,11 @@ Released in January 2026, [Ultralytics YOLO26](https://platform.ultralytics.com/ ### Why YOLO26 Outperforms the Competition -1. **End-to-End NMS-Free Design:** First pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates NMS post-processing. This delivers the deployment simplicity of RTDETRv2 while maintaining the lightning-fast speed of a highly optimized CNN. +1. **End-to-End NMS-Free Design:** First pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates NMS post-processing. This delivers the deployment simplicity of RTDETRv2 while maintaining the lightning-fast speed of a highly optimized CNN. 2. **MuSGD Optimizer:** Inspired by large language model innovations (such as Moonshot AI's Kimi K2), YOLO26 utilizes a hybrid of SGD and Muon. This ensures incredibly stable training dynamics and rapid convergence, reducing the time and compute resources required for custom datasets. 3. **Unmatched Edge Performance:** By executing complete DFL Removal (Distribution Focal Loss), YOLO26 simplifies export architectures. This optimization yields up to **43% faster CPU inference** compared to legacy models, making it the undisputed champion for edge AI and IoT devices. 4. **Enhanced Small Object Detection:** The introduction of ProgLoss and STAL loss functions provides a massive leap in detecting small objects—a critical requirement for drone analytics and aerial imagery that YOLOv6 historically struggled with. -5. **Task Versatility:** Unlike YOLOv6, which focuses strictly on detection, YOLO26 supports multi-modal workflows including [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/)—all from a single, unified API. +5. **Task Versatility:** Unlike YOLOv6, which focuses strictly on detection, YOLO26 supports multi-modal workflows including [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Image Classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb)—all from a single, unified API. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -175,4 +175,4 @@ model.export(format="onnx") Both YOLOv6-3.0 and RTDETRv2 are impressive contributions to the AI community. YOLOv6-3.0 remains a powerful tool for raw GPU industrial automation, and RTDETRv2 proves that transformer architectures can achieve real-time latency while maximizing accuracy. -However, for teams that require a reliable, production-ready framework with active community support, **Ultralytics YOLO models** are consistently the better choice. The seamless integration with platforms like [Hugging Face](https://huggingface.co/) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), combined with the incredibly low memory overhead during training, democratizes access to high-end AI. By upgrading to [YOLO26](https://docs.ultralytics.com/models/yolo26/), developers can leverage the groundbreaking MuSGD optimizer and NMS-free architecture to build faster, smarter, and more scalable computer vision pipelines. +However, for teams that require a reliable, production-ready framework with active community support, **Ultralytics YOLO models** are consistently the better choice. The seamless integration with platforms like [Hugging Face](https://huggingface.co/) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), combined with the incredibly low memory overhead during training, democratizes access to high-end AI. By upgrading to [YOLO26](https://docs.ultralytics.com/models/yolo26), developers can leverage the groundbreaking MuSGD optimizer and NMS-free architecture to build faster, smarter, and more scalable computer vision pipelines. diff --git a/docs/en/compare/yolov6-vs-yolo11.md b/docs/en/compare/yolov6-vs-yolo11.md index 415099f03ca..b1e24e8ecd3 100644 --- a/docs/en/compare/yolov6-vs-yolo11.md +++ b/docs/en/compare/yolov6-vs-yolo11.md @@ -17,40 +17,40 @@ Both models offer strong solutions for [machine learning](https://www.ultralytic ## YOLOv6-3.0: Industrial Throughput Specialization -Developed by the Vision AI Department at Meituan, YOLOv6-3.0 is positioned as a next-generation [object detection](https://docs.ultralytics.com/tasks/detect/) framework explicitly optimized for industrial applications. +Developed by the Vision AI Department at Meituan, YOLOv6-3.0 is positioned as a next-generation [object detection](https://docs.ultralytics.com/tasks/detect) framework explicitly optimized for industrial applications. - **Authors:** Chuyi Li, Lulu Li, Yifei Geng, Hongliang Jiang, Meng Cheng, Bo Zhang, Zaidan Ke, Xiaoming Xu, and Xiangxiang Chu - **Organization:** [Meituan](https://tech.meituan.com/) - **Date:** 2023-01-13 - **Arxiv:** [2301.05586](https://arxiv.org/abs/2301.05586) - **GitHub:** [meituan/YOLOv6](https://github.com/meituan/YOLOv6) -- **Docs:** [YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6/) +- **Docs:** [YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6) ### Architecture Highlights -YOLOv6-3.0 focuses heavily on maximizing throughput on hardware accelerators like NVIDIA GPUs. Its backbone relies on an **EfficientRep** design, which is highly hardware-friendly for GPU inference operations using platforms like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). +YOLOv6-3.0 focuses heavily on maximizing throughput on hardware accelerators like NVIDIA GPUs. Its backbone relies on an **EfficientRep** design, which is highly hardware-friendly for GPU inference operations using platforms like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). A major architectural feature is the **Bi-directional Concatenation (BiC)** module in its neck, which enhances feature fusion across different scales. To improve convergence during the training phase, YOLOv6 employs an **Anchor-Aided Training (AAT)** strategy. This strategy temporarily leverages [anchor boxes](https://www.ultralytics.com/glossary/anchor-boxes) during training to reap the benefits of anchor-based paradigms, while inference fundamentally remains anchor-free. While YOLOv6-3.0 excels in high-speed, batch-processing environments such as offline video analytics on powerful server-grade hardware, this deep specialization can sometimes result in sub-optimal latency on CPU-only edge devices compared to models designed for broader general-purpose computing. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Ultralytics YOLO11: The Versatile Multi-Task Standard -Released by Ultralytics, [YOLO11](https://docs.ultralytics.com/models/yolo11/) represents a major shift toward a unified, highly efficient framework capable of handling a massive array of vision tasks simultaneously. +Released by Ultralytics, [YOLO11](https://docs.ultralytics.com/models/yolo11) represents a major shift toward a unified, highly efficient framework capable of handling a massive array of vision tasks simultaneously. - **Authors:** Glenn Jocher and Jing Qiu - **Organization:** [Ultralytics](https://www.ultralytics.com/about) - **Date:** 2024-09-27 - **GitHub:** [ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11/) +- **Docs:** [YOLO11 Documentation](https://docs.ultralytics.com/models/yolo11) ### The Ultralytics Advantage While specialized industrial models are valuable, most modern developers prioritize a balance of performance, ease of use, memory efficiency, and diverse task support. YOLO11 shines by providing a comprehensive solution. -Unlike YOLOv6, which focuses strictly on bounding box detection, Ultralytics YOLO11 is natively equipped for [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) extraction. It achieves this while maintaining an incredibly accessible ecosystem. +Unlike YOLOv6, which focuses strictly on bounding box detection, Ultralytics YOLO11 is natively equipped for [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) extraction. It achieves this while maintaining an incredibly accessible ecosystem. !!! tip "Streamlined Machine Learning Workflows" @@ -77,7 +77,7 @@ The table below provides a detailed look at how these models perform across diff ### Memory Requirements and Training Efficiency -When preparing custom data, training efficiency is paramount. Ultralytics YOLO models require significantly lower VRAM usage during training than heavily customized industrial networks or massive transformer-based architectures. This democratizes AI, allowing researchers to fine-tune high-accuracy models on consumer-grade GPUs. Furthermore, the active Ultralytics community ensures that tools like [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/) and logging integrations (like Weights & Biases or [Comet ML](https://docs.ultralytics.com/integrations/comet/)) are always up to date. +When preparing custom data, training efficiency is paramount. Ultralytics YOLO models require significantly lower VRAM usage during training than heavily customized industrial networks or massive transformer-based architectures. This democratizes AI, allowing researchers to fine-tune high-accuracy models on consumer-grade GPUs. Furthermore, the active Ultralytics community ensures that tools like [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning) and logging integrations (like Weights & Biases or [Comet ML](https://docs.ultralytics.com/integrations/comet)) are always up to date. ## Use Cases and Recommendations @@ -95,21 +95,21 @@ YOLOv6 is a strong choice for: YOLO11 is recommended for: -- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) where reliability and active maintenance are paramount. -- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [OBB](https://docs.ultralytics.com/tasks/obb/) within a single unified framework. -- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python/). +- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) where reliability and active maintenance are paramount. +- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [OBB](https://docs.ultralytics.com/tasks/obb) within a single unified framework. +- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python). ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Code Example: The Unified Python API -Training a state-of-the-art model with Ultralytics takes only a few lines of code. This same API handles predictions, validations, and exports to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino/). +Training a state-of-the-art model with Ultralytics takes only a few lines of code. This same API handles predictions, validations, and exports to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino). ```python from ultralytics import YOLO @@ -133,7 +133,7 @@ While YOLO11 stands tall as a massive leap over legacy architectures, developers Released in January 2026, YOLO26 establishes a new standard for AI model efficiency, bringing innovations previously unseen in the computer vision space: -- **End-to-End NMS-Free Design:** Bypassing the need for [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) reduces deployment latency drastically—a method first introduced in [YOLOv10](https://docs.ultralytics.com/models/yolov10/). +- **End-to-End NMS-Free Design:** Bypassing the need for [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) reduces deployment latency drastically—a method first introduced in [YOLOv10](https://docs.ultralytics.com/models/yolov10). - **MuSGD Optimizer:** Integrating LLM training stability into vision tasks, this optimizer combines SGD and Muon for incredibly stable and fast convergence. - **CPU Optimized:** By removing the Distribution Focal Loss (DFL), YOLO26 achieves up to 43% faster CPU inference, making it the perfect choice for mobile, IoT, and [edge AI applications](https://www.ultralytics.com/glossary/edge-ai). - **Advanced Loss Functions:** Implementations of ProgLoss and STAL drastically improve small-object recognition, vital for aerial imagery and robotics. @@ -144,4 +144,4 @@ Released in January 2026, YOLO26 establishes a new standard for AI model efficie If your deployment environment is strictly confined to heavily engineered industrial GPU pipelines requiring batch inference, **YOLOv6-3.0** remains an interesting tool. However, for the vast majority of real-world scenarios requiring scalable, easy-to-train, and highly accurate models, **Ultralytics YOLO11**—and the cutting-edge **YOLO26**—are the undisputed recommendations. -The Ultralytics ecosystem empowers you to move rapidly from dataset collection to edge deployment, ensuring your projects are future-proof and backed by extensive documentation and community support. For those exploring other efficient architectures, we also recommend checking out [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) for robust, proven legacy support, or dive directly into the next generation with [YOLO26](https://docs.ultralytics.com/models/yolo26/). +The Ultralytics ecosystem empowers you to move rapidly from dataset collection to edge deployment, ensuring your projects are future-proof and backed by extensive documentation and community support. For those exploring other efficient architectures, we also recommend checking out [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) for robust, proven legacy support, or dive directly into the next generation with [YOLO26](https://docs.ultralytics.com/models/yolo26). diff --git a/docs/en/compare/yolov6-vs-yolo26.md b/docs/en/compare/yolov6-vs-yolo26.md index 87d2b7f85c6..87cbdc6a9ad 100644 --- a/docs/en/compare/yolov6-vs-yolo26.md +++ b/docs/en/compare/yolov6-vs-yolo26.md @@ -6,7 +6,7 @@ keywords: YOLOv6-3.0, YOLO26, object detection, model comparison, computer visio # YOLOv6-3.0 vs YOLO26: A Deep Dive into Real-Time Object Detection -The evolution of real-time [object detection](https://docs.ultralytics.com/tasks/detect/) has brought forth incredible innovations, often polarizing the focus between industrial GPU throughput and versatile, edge-optimized architectures. In this comprehensive comparison, we explore the nuances between two heavyweights: the industrially focused **YOLOv6-3.0** and the newly released, natively end-to-end **Ultralytics YOLO26**. +The evolution of real-time [object detection](https://docs.ultralytics.com/tasks/detect) has brought forth incredible innovations, often polarizing the focus between industrial GPU throughput and versatile, edge-optimized architectures. In this comprehensive comparison, we explore the nuances between two heavyweights: the industrially focused **YOLOv6-3.0** and the newly released, natively end-to-end **Ultralytics YOLO26**. Whether you are deploying to high-end server GPUs or low-power edge devices, understanding the architectural strengths and ideal use cases of these models is crucial for optimizing your computer vision pipelines. @@ -24,13 +24,13 @@ Developed by the Meituan Vision AI Department, YOLOv6-3.0 was designed as a "nex - **Date:** 2023-01-13 - **Arxiv:** [2301.05586](https://arxiv.org/abs/2301.05586) - **GitHub:** [meituan/YOLOv6](https://github.com/meituan/YOLOv6) -- **Docs:** [YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6/) +- **Docs:** [YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6) ### Architectural Focus YOLOv6-3.0 employs a **Bi-directional Concatenation (BiC)** module in its neck to improve feature fusion, combined with an **Anchor-Aided Training (AAT)** strategy. Its backbone is based on **EfficientRep**, a topology engineered to be highly hardware-friendly for GPU inference. While this makes it exceptionally fast when leveraging [NVIDIA TensorRT](https://developer.nvidia.com/tensorrt), it can lead to higher latency on CPU-only or edge devices that lack massive parallel processing capabilities. -[Learn more about YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6){ .md-button } ## YOLO26: The New Standard for Edge and Cloud @@ -46,17 +46,17 @@ Released in January 2026, **Ultralytics YOLO26** represents a paradigm shift. It YOLO26 introduces several pioneering advancements that set it apart from previous generations: -- **End-to-End NMS-Free Design:** Building on concepts first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 is natively end-to-end. It completely eliminates [Non-Maximum Suppression (NMS)](https://en.wikipedia.org/wiki/NMS) post-processing, resulting in a dramatic reduction in latency variability and drastically simpler deployment logic. +- **End-to-End NMS-Free Design:** Building on concepts first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 is natively end-to-end. It completely eliminates [Non-Maximum Suppression (NMS)](https://en.wikipedia.org/wiki/NMS) post-processing, resulting in a dramatic reduction in latency variability and drastically simpler deployment logic. - **Up to 43% Faster CPU Inference:** Optimized explicitly for edge computing, YOLO26 excels on devices without GPUs, making it ideal for mobile phones, IoT sensors, and robotics. - **DFL Removal:** The Distribution Focal Loss has been removed, simplifying the model export process and enhancing compatibility with low-power edge devices. - **MuSGD Optimizer:** Inspired by LLM training innovations like Moonshot AI's Kimi K2, the new MuSGD optimizer (a hybrid of [Stochastic Gradient Descent](https://en.wikipedia.org/wiki/Stochastic_gradient_descent) and Muon) brings large-scale stability to vision tasks, ensuring faster convergence. -- **ProgLoss + STAL:** Advanced loss functions yield notable improvements in small-object recognition, a critical enhancement for applications dealing with [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and crowded scenes. +- **ProgLoss + STAL:** Advanced loss functions yield notable improvements in small-object recognition, a critical enhancement for applications dealing with [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and crowded scenes. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } !!! tip "Multi-Task Capabilities" - Unlike YOLOv6-3.0, which strictly handles bounding boxes, YOLO26 features task-specific improvements across the board. This includes semantic segmentation loss and multi-scale proto for [instance segmentation](https://docs.ultralytics.com/tasks/segment/), Residual Log-Likelihood Estimation (RLE) for [pose estimation](https://docs.ultralytics.com/tasks/pose/), and specialized angle loss to resolve [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) boundary issues. + Unlike YOLOv6-3.0, which strictly handles bounding boxes, YOLO26 features task-specific improvements across the board. This includes semantic segmentation loss and multi-scale proto for [instance segmentation](https://docs.ultralytics.com/tasks/segment), Residual Log-Likelihood Estimation (RLE) for [pose estimation](https://docs.ultralytics.com/tasks/pose), and specialized angle loss to resolve [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) boundary issues. ## Detailed Performance Comparison @@ -82,9 +82,9 @@ As seen in the data, YOLO26 consistently achieves a superior **Performance Balan Choosing a model involves evaluating the surrounding software ecosystem. Here, the Ultralytics suite provides decisive benefits over static research repositories: - **Ease of Use:** Ultralytics provides a "zero-to-hero" developer experience. Its unified Python API allows users to switch between tasks and models simply by altering a single string parameter. -- **Well-Maintained Ecosystem:** Through the [Ultralytics Platform](https://platform.ultralytics.com), developers gain access to an actively updated environment that supports continuous dataset management, cloud training, and seamless [model export](https://docs.ultralytics.com/modes/export/) to formats like [ONNX](https://onnx.ai/) and OpenVINO. -- **Memory Requirements:** YOLO26 boasts a highly efficient training methodology with significantly lower memory requirements during both training and inference. This contrasts favorably against transformer-based architectures, such as [RT-DETR](https://docs.ultralytics.com/models/rtdetr/), which demand massive CUDA memory allocations. -- **Versatility:** By natively supporting [classification](https://docs.ultralytics.com/tasks/classify/), detection, segmentation, and pose estimation, YOLO26 serves as a one-stop-shop for complex, multi-modal vision applications. +- **Well-Maintained Ecosystem:** Through the [Ultralytics Platform](https://platform.ultralytics.com), developers gain access to an actively updated environment that supports continuous dataset management, cloud training, and seamless [model export](https://docs.ultralytics.com/modes/export) to formats like [ONNX](https://onnx.ai/) and OpenVINO. +- **Memory Requirements:** YOLO26 boasts a highly efficient training methodology with significantly lower memory requirements during both training and inference. This contrasts favorably against transformer-based architectures, such as [RT-DETR](https://docs.ultralytics.com/models/rtdetr), which demand massive CUDA memory allocations. +- **Versatility:** By natively supporting [classification](https://docs.ultralytics.com/tasks/classify), detection, segmentation, and pose estimation, YOLO26 serves as a one-stop-shop for complex, multi-modal vision applications. !!! note "Exploring Alternatives" diff --git a/docs/en/compare/yolov6-vs-yolov10.md b/docs/en/compare/yolov6-vs-yolov10.md index f5d40068aa0..c61d42dd924 100644 --- a/docs/en/compare/yolov6-vs-yolov10.md +++ b/docs/en/compare/yolov6-vs-yolov10.md @@ -29,7 +29,7 @@ The core of YOLOv6-3.0 lies in its hardware-friendly design. It incorporates a B Powered by an EfficientRep backbone, this model shines in heavy-duty [manufacturing automation](https://www.ultralytics.com/blog/manufacturing-automation) tasks where batch processing on powerful NVIDIA hardware (such as T4 or A100 GPUs) is the norm. While it performs admirably in server clusters, its reliance on specific hardware optimizations can make it less efficient on low-power edge CPUs. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Unpacking YOLOv10: The NMS-Free Pioneer @@ -47,7 +47,7 @@ YOLOv10's major contribution to the field is its end-to-end NMS-free design. By Furthermore, the model boasts a holistic efficiency-accuracy driven model design. Through comprehensive optimization of various layers, YOLOv10 drastically cuts down computational redundancy. This makes it highly suitable for resource-constrained environments, including [autonomous vehicles](https://www.ultralytics.com/glossary/autonomous-vehicles) and edge robotics. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## Detailed Performance Comparison @@ -97,11 +97,11 @@ YOLOv10 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Why YOLO26 is the Superior Choice @@ -114,18 +114,18 @@ Released in January 2026, YOLO26 incorporates the best innovations from the prec - **End-to-End NMS-Free Design:** Building on the concept pioneered in YOLOv10, YOLO26 natively eliminates NMS post-processing, resulting in smoother, more predictable inference times that are drastically [easier to ship to production](https://www.ultralytics.com/blog/exploring-why-ultralytics-yolo26-is-easier-to-ship-to-production). - **MuSGD Optimizer:** Inspired by large language model optimizations like Moonshot AI's Kimi K2, this hybrid of SGD and Muon ensures incredibly stable training and dramatically faster convergence. - **Up to 43% Faster CPU Inference:** For edge devices, YOLO26 features specific architectural simplifications, making it vastly superior for deployment on IoT chips and consumer CPUs. -- **DFL Removal:** The removal of Distribution Focal Loss simplifies the head export, greatly improving compatibility with low-power deployment engines like [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) or NCNN. +- **DFL Removal:** The removal of Distribution Focal Loss simplifies the head export, greatly improving compatibility with low-power deployment engines like [OpenVINO](https://docs.ultralytics.com/integrations/openvino) or NCNN. - **ProgLoss + STAL:** Advanced loss formulations notably boost precision on small object recognition, which is critical for [drone UAV operations](https://www.ultralytics.com/blog/computer-vision-applications-ai-drone-uav-operations) and distant subject tracking. -Furthermore, unlike single-task repositories, the Ultralytics ecosystem handles a massive array of vision tasks out of the box, including bounding box detection, [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/). +Furthermore, unlike single-task repositories, the Ultralytics ecosystem handles a massive array of vision tasks out of the box, including bounding box detection, [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose). ### Training Efficiency and Memory Optimization -A critical advantage of Ultralytics YOLO models over complex [transformer-based architectures like RT-DETR](https://docs.ultralytics.com/models/rtdetr/) is their incredibly low CUDA memory consumption during training. A developer can comfortably fine-tune YOLO26 on a consumer-grade GPU or through free cloud resources, significantly democratizing AI development. +A critical advantage of Ultralytics YOLO models over complex [transformer-based architectures like RT-DETR](https://docs.ultralytics.com/models/rtdetr) is their incredibly low CUDA memory consumption during training. A developer can comfortably fine-tune YOLO26 on a consumer-grade GPU or through free cloud resources, significantly democratizing AI development. ### Code Example: Getting Started with YOLO26 -The ease of use provided by the [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) allows you to load, train, and test models in just a few lines of code. +The ease of use provided by the [Ultralytics Python API](https://docs.ultralytics.com/usage/python) allows you to load, train, and test models in just a few lines of code. ```python from ultralytics import YOLO @@ -154,4 +154,4 @@ When choosing between YOLOv6-3.0 and YOLOv10, the decision hinges on the deploym However, for developers seeking zero-compromise performance backed by comprehensive documentation, cloud logging via the [Ultralytics Platform](https://platform.ultralytics.com), and multi-task versatility, **YOLO26 is the definitive recommendation**. -For legacy infrastructure requirements, teams might also investigate the previous generation [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), or explore [YOLO-World](https://docs.ultralytics.com/models/yolo-world/) for unique open-vocabulary detection capabilities. +For legacy infrastructure requirements, teams might also investigate the previous generation [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), or explore [YOLO-World](https://docs.ultralytics.com/models/yolo-world) for unique open-vocabulary detection capabilities. diff --git a/docs/en/compare/yolov6-vs-yolov5.md b/docs/en/compare/yolov6-vs-yolov5.md index 8625a38ec02..03f2f05ca12 100644 --- a/docs/en/compare/yolov6-vs-yolov5.md +++ b/docs/en/compare/yolov6-vs-yolov5.md @@ -22,7 +22,7 @@ Developed by the Vision AI Department at [Meituan](https://en.wikipedia.org/wiki - **Date:** 2023-01-13 - **Arxiv:** [2301.05586](https://arxiv.org/abs/2301.05586) - **GitHub:** [meituan/YOLOv6](https://github.com/meituan/YOLOv6) -- **Docs:** [YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6/) +- **Docs:** [YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6) ### Architectural Strengths @@ -30,7 +30,7 @@ YOLOv6-3.0 introduces several structural optimizations designed for speed. The m During the training phase, the model incorporates an **Anchor-Aided Training (AAT)** strategy. This approach attempts to marry the stability of anchor-based training with the speed of anchor-free inference. Additionally, its neck architecture uses a **Bi-directional Concatenation (BiC)** module to improve feature fusion across different scales. While highly optimized for high-end server GPUs using [TensorRT](https://developer.nvidia.com/tensorrt), this specialization can sometimes result in increased latency on CPU-only or low-power edge devices. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Ultralytics YOLOv5: The Pioneer of Accessible Vision AI @@ -46,7 +46,7 @@ Released by Ultralytics, YOLOv5 set a new standard for ease of use, training eff The defining characteristic of YOLOv5 is its **Ease of Use**. Built natively on the [PyTorch](https://pytorch.org/) framework, the repository provides a unified Python API that drastically simplifies the [machine learning](https://www.ultralytics.com/glossary/machine-learning-ml) lifecycle. From dataset configuration to final deployment, the integrated ecosystem ensures that developers spend less time debugging environments and more time building applications. -YOLOv5 is not just limited to [object detection](https://docs.ultralytics.com/tasks/detect/). It boasts exceptional **Versatility**, natively supporting [image classification](https://docs.ultralytics.com/tasks/classify/) and [instance segmentation](https://docs.ultralytics.com/tasks/segment/). Furthermore, it offers unparalleled **Training Efficiency**, featuring smart caching, automated data loaders, and built-in support for distributed multi-GPU training. +YOLOv5 is not just limited to [object detection](https://docs.ultralytics.com/tasks/detect). It boasts exceptional **Versatility**, natively supporting [image classification](https://docs.ultralytics.com/tasks/classify) and [instance segmentation](https://docs.ultralytics.com/tasks/segment). Furthermore, it offers unparalleled **Training Efficiency**, featuring smart caching, automated data loaders, and built-in support for distributed multi-GPU training. !!! tip "Memory Efficiency in Ultralytics Models" @@ -101,7 +101,7 @@ model.export(format="onnx") Choosing between these architectures often depends on your specific infrastructure constraints: - **When to deploy YOLOv6-3.0:** Ideal for automated manufacturing lines and high-throughput server analytics where dedicated NVIDIA GPUs are available and latency must be minimal. Its architecture thrives in environments where TensorRT optimizations can be fully utilized. -- **When to deploy YOLOv5:** The perfect choice for rapid prototyping, cross-platform deployment, and teams looking for a unified pipeline. Its diverse export capabilities make it ideal for retail analytics on edge devices, agricultural drone monitoring, and [pose estimation](https://docs.ultralytics.com/tasks/pose/) in fitness applications. +- **When to deploy YOLOv5:** The perfect choice for rapid prototyping, cross-platform deployment, and teams looking for a unified pipeline. Its diverse export capabilities make it ideal for retail analytics on edge devices, agricultural drone monitoring, and [pose estimation](https://docs.ultralytics.com/tasks/pose) in fitness applications. ## The Future of Object Detection: Enter YOLO26 diff --git a/docs/en/compare/yolov6-vs-yolov7.md b/docs/en/compare/yolov6-vs-yolov7.md index 7f1d5090b70..da702f265d7 100644 --- a/docs/en/compare/yolov6-vs-yolov7.md +++ b/docs/en/compare/yolov6-vs-yolov7.md @@ -31,7 +31,7 @@ YOLOv6-3.0 relies on an **EfficientRep** backbone, a hardware-friendly architect Furthermore, YOLOv6-3.0 implements an **Anchor-Aided Training (AAT)** strategy. This approach combines the rich gradient signals of anchor-based training with the streamlined deployment benefits of anchor-free inference, helping the model converge more stably without sacrificing post-processing speed. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } !!! info "Hardware Considerations" @@ -53,7 +53,7 @@ The core of YOLOv7 is its **Extended Efficient Layer Aggregation Network (E-ELAN YOLOv7 also heavily utilizes model re-parameterization, merging convolutional layers with batch normalization during inference. This reduces the parameter count and speeds up the forward pass when deployed using frameworks like [NVIDIA TensorRT](https://developer.nvidia.com/tensorrt) or [ONNX](https://onnx.ai/). -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ## Performance Comparison @@ -93,33 +93,33 @@ YOLOv7 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Stepping into the Future -While YOLOv6-3.0 and YOLOv7 represent significant milestones, integrating disparate repositories into production pipelines often presents challenges in [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/) and hyperparameter tuning. The **Ultralytics ecosystem** resolves these pain points by offering a streamlined, unified interface. +While YOLOv6-3.0 and YOLOv7 represent significant milestones, integrating disparate repositories into production pipelines often presents challenges in [model deployment](https://docs.ultralytics.com/guides/model-deployment-options) and hyperparameter tuning. The **Ultralytics ecosystem** resolves these pain points by offering a streamlined, unified interface. ### Why Choose Ultralytics? - **Ease of Use:** The Ultralytics Python API allows developers to load, train, and export models with just a few lines of code. Switching from an older model to the latest architecture requires changing only a single string. - **Well-Maintained Ecosystem:** Ultralytics provides frequent updates, active community support, and robust [documentation](https://docs.ultralytics.com/). -- **Versatility:** Unlike earlier models that focused primarily on bounding boxes, Ultralytics models natively support multi-task learning, including [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). -- **Memory Requirements:** Ultralytics YOLO models maintain lower memory usage during training compared to transformer-based architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/), allowing researchers to train effectively on consumer-grade hardware. +- **Versatility:** Unlike earlier models that focused primarily on bounding boxes, Ultralytics models natively support multi-task learning, including [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb). +- **Memory Requirements:** Ultralytics YOLO models maintain lower memory usage during training compared to transformer-based architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr), allowing researchers to train effectively on consumer-grade hardware. ### Upgrading to YOLO26 -For developers seeking the pinnacle of performance, **YOLO26** (released January 2026) fundamentally shifts the paradigm of [object detection](https://docs.ultralytics.com/tasks/detect/). It introduces a fully **End-to-End NMS-Free Design**, eliminating complex post-processing logic and severely reducing latency variance on edge devices. +For developers seeking the pinnacle of performance, **YOLO26** (released January 2026) fundamentally shifts the paradigm of [object detection](https://docs.ultralytics.com/tasks/detect). It introduces a fully **End-to-End NMS-Free Design**, eliminating complex post-processing logic and severely reducing latency variance on edge devices. Key innovations in YOLO26 include: - **MuSGD Optimizer:** A sophisticated hybrid of SGD and Muon that ensures incredibly stable training dynamics and faster convergence. - **DFL Removal:** By stripping out Distribution Focal Loss, YOLO26 simplifies export compatibility and boosts performance on low-power devices. - **ProgLoss + STAL:** Advanced loss functions that yield notable improvements in small-object recognition. -- **Unrivaled Speed:** Achieves up to 43% faster CPU inference compared to previous generations, making it perfect for embedded systems like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [Apple CoreML](https://docs.ultralytics.com/integrations/coreml/) deployments. +- **Unrivaled Speed:** Achieves up to 43% faster CPU inference compared to previous generations, making it perfect for embedded systems like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [Apple CoreML](https://docs.ultralytics.com/integrations/coreml) deployments. Other highly capable models within the ecosystem include [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), both of which offer excellent performance balance for legacy hardware integrations. diff --git a/docs/en/compare/yolov6-vs-yolov8.md b/docs/en/compare/yolov6-vs-yolov8.md index ca67ffb02dc..86d7a3aea46 100644 --- a/docs/en/compare/yolov6-vs-yolov8.md +++ b/docs/en/compare/yolov6-vs-yolov8.md @@ -22,7 +22,7 @@ Developed by the Vision AI Department at [Meituan](https://en.wikipedia.org/wiki - **Date:** 2023-01-13 - **Arxiv:** [2301.05586](https://arxiv.org/abs/2301.05586) - **GitHub:** [meituan/YOLOv6](https://github.com/meituan/YOLOv6) -- **Docs:** [Ultralytics YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6/) +- **Docs:** [Ultralytics YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6) ### Architectural Focus @@ -30,7 +30,7 @@ YOLOv6-3.0 leverages an **EfficientRep** backbone, a hardware-friendly architect During the training phase, YOLOv6 incorporates an Anchor-Aided Training (AAT) strategy. This hybrid approach attempts to capture the benefits of both anchor-based and anchor-free paradigms while maintaining an anchor-free inference pipeline. While highly effective for dedicated [TensorRT](https://developer.nvidia.com/tensorrt) deployments, this specialization can result in higher latency on CPU-only edge devices. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Ultralytics YOLOv8: The Versatile Multi-Task Standard @@ -44,9 +44,9 @@ Released by Ultralytics, YOLOv8 represents a paradigm shift from specialized bou ### Architectural Highlights -YOLOv8 natively features a decoupled head structure that separates objectness, classification, and regression tasks, significantly improving convergence speed. Its anchor-free design eliminates the need for manual anchor box configuration, ensuring robust generalization across highly diverse [computer vision datasets](https://docs.ultralytics.com/datasets/detect/). +YOLOv8 natively features a decoupled head structure that separates objectness, classification, and regression tasks, significantly improving convergence speed. Its anchor-free design eliminates the need for manual anchor box configuration, ensuring robust generalization across highly diverse [computer vision datasets](https://docs.ultralytics.com/datasets/detect). -The model integrates the advanced **C2f module** (Cross-Stage Partial bottleneck with two convolutions), replacing older C3 blocks. This enhances gradient flow and feature representation without inflating the computational budget. Crucially, YOLOv8 is not just a detection engine; it natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) tasks within a single API. +The model integrates the advanced **C2f module** (Cross-Stage Partial bottleneck with two convolutions), replacing older C3 blocks. This enhances gradient flow and feature representation without inflating the computational budget. Crucially, YOLOv8 is not just a detection engine; it natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) tasks within a single API. [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -87,33 +87,33 @@ YOLOv6 is a strong choice for: YOLOv8 is recommended for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Ecosystem and Ease of Use While raw inference speed is important, the lifecycle of a machine learning project involves data management, training, exporting, and monitoring. The integrated [Ultralytics Platform](https://platform.ultralytics.com/) provides a seamless "zero-to-hero" experience that research-only repositories struggle to match. - **Well-Maintained Ecosystem:** Ultralytics provides frequent updates, ensuring compatibility with the latest [PyTorch](https://pytorch.org/) releases and hardware drivers. -- **Ease of Use:** A unified Python API allows developers to train and export models to formats like [ONNX](https://onnx.ai/) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) with a single line of code. -- **Lower Memory Requirements:** Ultralytics models are highly optimized to minimize CUDA memory usage during training, making advanced AI accessible on consumer-grade hardware—a stark contrast to memory-hungry transformer architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +- **Ease of Use:** A unified Python API allows developers to train and export models to formats like [ONNX](https://onnx.ai/) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino) with a single line of code. +- **Lower Memory Requirements:** Ultralytics models are highly optimized to minimize CUDA memory usage during training, making advanced AI accessible on consumer-grade hardware—a stark contrast to memory-hungry transformer architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). ## Looking Forward: The Ultimate Upgrade to YOLO26 -For developers seeking the pinnacle of performance and modern deployment capabilities, [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) (released January 2026) is the recommended standard. It builds upon the successes of YOLOv8 and the previous [YOLO11](https://docs.ultralytics.com/models/yolo11/) generation, introducing revolutionary architectural improvements: +For developers seeking the pinnacle of performance and modern deployment capabilities, [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) (released January 2026) is the recommended standard. It builds upon the successes of YOLOv8 and the previous [YOLO11](https://docs.ultralytics.com/models/yolo11) generation, introducing revolutionary architectural improvements: -- **End-to-End NMS-Free Design:** YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing, a concept pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/). This streamlines deployment logic and reduces latency variance. +- **End-to-End NMS-Free Design:** YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing, a concept pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10). This streamlines deployment logic and reduces latency variance. - **MuSGD Optimizer:** Inspired by large language model innovations like Moonshot AI's Kimi K2, the new MuSGD optimizer (a hybrid of SGD and Muon) stabilizes training and accelerates convergence across diverse datasets. -- **DFL Removal & CPU Speed:** By removing Distribution Focal Loss (DFL), YOLO26 simplifies its export graph. This optimization unlocks **up to 43% faster CPU inference**, making it the absolute best choice for [mobile and IoT edge computing](https://docs.ultralytics.com/guides/model-deployment-options/). +- **DFL Removal & CPU Speed:** By removing Distribution Focal Loss (DFL), YOLO26 simplifies its export graph. This optimization unlocks **up to 43% faster CPU inference**, making it the absolute best choice for [mobile and IoT edge computing](https://docs.ultralytics.com/guides/model-deployment-options). - **ProgLoss + STAL:** Advanced loss functions deliver notable improvements in small-object recognition, which is critical for aerial drone imagery and robotics. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -148,4 +148,4 @@ print(f"Model exported successfully to: {export_path}") Choosing the right architecture dictates the long-term maintainability of your pipeline. **YOLOv6-3.0** serves as a specialized tool for industrial pipelines with heavy GPU accelerators. However, **Ultralytics YOLOv8** provides a superior balance of multi-task versatility, lower parameter counts, and an unmatched training ecosystem. -For new implementations, upgrading to **YOLO26** via the [Ultralytics Platform](https://platform.ultralytics.com) ensures you are utilizing the absolute fastest, natively end-to-end, NMS-free architecture available today, future-proofing your [AI deployment strategies](https://docs.ultralytics.com/guides/model-deployment-practices/). +For new implementations, upgrading to **YOLO26** via the [Ultralytics Platform](https://platform.ultralytics.com) ensures you are utilizing the absolute fastest, natively end-to-end, NMS-free architecture available today, future-proofing your [AI deployment strategies](https://docs.ultralytics.com/guides/model-deployment-practices). diff --git a/docs/en/compare/yolov6-vs-yolov9.md b/docs/en/compare/yolov6-vs-yolov9.md index 1d429831831..a9d111c30ab 100644 --- a/docs/en/compare/yolov6-vs-yolov9.md +++ b/docs/en/compare/yolov6-vs-yolov9.md @@ -17,7 +17,7 @@ While both models offer unique architectural innovations, developers looking for ## YOLOv6-3.0: Industrial Throughput Optimization -Developed by the Vision AI Department at Meituan, [YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6/) was heavily engineered for maximum throughput in industrial applications, particularly on GPU hardware. +Developed by the Vision AI Department at Meituan, [YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6) was heavily engineered for maximum throughput in industrial applications, particularly on GPU hardware. - **Authors:** Chuyi Li, Lulu Li, Yifei Geng, Hongliang Jiang, Meng Cheng, Bo Zhang, Zaidan Ke, Xiaoming Xu, and Xiangxiang Chu - **Organization:** [Meituan](https://github.com/meituan) @@ -35,11 +35,11 @@ The backbone is based on an EfficientRep design, meticulously optimized to be ha The primary strength of YOLOv6-3.0 lies in its high frame rate on GPUs like the NVIDIA T4, making it suitable for high-density [video understanding](https://www.ultralytics.com/glossary/video-understanding) streams. However, its heavy reliance on specific hardware optimizations can result in sub-optimal latency on CPU-only edge devices. Furthermore, setting up its training pipeline can be complex compared to more unified frameworks. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## YOLOv9: Programmable Gradient Information -Released a year later, [YOLOv9](https://docs.ultralytics.com/models/yolov9/) focuses on solving the information bottleneck problem inherent in deep neural networks, pushing the theoretical limits of CNN architectures. +Released a year later, [YOLOv9](https://docs.ultralytics.com/models/yolov9) focuses on solving the information bottleneck problem inherent in deep neural networks, pushing the theoretical limits of CNN architectures. - **Authors:** Chien-Yao Wang and Hong-Yuan Mark Liao - **Organization:** [Institute of Information Science, Academia Sinica](https://www.iis.sinica.edu.tw/zh/index.html) @@ -55,7 +55,7 @@ YOLOv9's major contribution is **Programmable Gradient Information (PGI)**, whic YOLOv9 achieves outstanding [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map) on benchmark datasets like COCO, making it a favorite for researchers prioritizing raw accuracy. However, like YOLOv6, it still relies on traditional Non-Maximum Suppression (NMS) for post-processing. This adds latency and complicates the [model deployment](https://www.ultralytics.com/glossary/model-deployment) pipeline, especially when porting to edge devices using formats like ONNX or TensorRT. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ## Performance Comparison @@ -89,13 +89,13 @@ Released in early 2026, YOLO26 fundamentally redefines deployment efficiency by 1. **MuSGD Optimizer:** Inspired by LLM training (like Moonshot AI's Kimi K2), YOLO26 utilizes a hybrid of SGD and Muon. This brings unparalleled training stability and faster convergence to computer vision tasks. 2. **Up to 43% Faster CPU Inference:** Unlike YOLOv6's heavy GPU focus, YOLO26 is heavily optimized for edge devices. The removal of Distribution Focal Loss (DFL) simplifies the head, making it highly compatible with low-power CPUs and [edge computing](https://www.ultralytics.com/glossary/edge-computing) hardware. 3. **ProgLoss + STAL:** Advanced loss functions drastically improve small object detection, which is critical for [aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision) and robotics. -4. **Unmatched Versatility:** While YOLOv6 is purely a detection engine, YOLO26 handles [instance segmentation](https://docs.ultralytics.com/tasks/segment/), classification, [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection seamlessly. +4. **Unmatched Versatility:** While YOLOv6 is purely a detection engine, YOLO26 handles [instance segmentation](https://docs.ultralytics.com/tasks/segment), classification, [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection seamlessly. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } ### Seamless Training with Ultralytics -Training state-of-the-art models shouldn't require complex bash scripts. The Ultralytics Python API provides a streamlined experience with automatic data loading, minimal [CUDA memory usage](https://docs.ultralytics.com/guides/yolo-common-issues/), and built-in tracking. +Training state-of-the-art models shouldn't require complex bash scripts. The Ultralytics Python API provides a streamlined experience with automatic data loading, minimal [CUDA memory usage](https://docs.ultralytics.com/guides/yolo-common-issues), and built-in tracking. ```python from ultralytics import YOLO @@ -116,6 +116,6 @@ Choosing the right architecture depends entirely on your target deployment envir - **Use YOLOv6-3.0 for:** Factory automation and defect detection where server-grade GPUs (e.g., A100s) are abundant and batch processing maximizes throughput. - **Use YOLOv9 for:** Academic research or competitions where wringing out the absolute highest mAP on standardized datasets like COCO is the primary goal. -- **Use YOLO26 for:** Almost all modern commercial applications. Its NMS-free architecture, low memory footprint, and high-speed CPU inference make it perfect for [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/), smart retail, and real-time [object tracking](https://docs.ultralytics.com/modes/track/) on embedded devices. +- **Use YOLO26 for:** Almost all modern commercial applications. Its NMS-free architecture, low memory footprint, and high-speed CPU inference make it perfect for [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system), smart retail, and real-time [object tracking](https://docs.ultralytics.com/modes/track) on embedded devices. By leveraging the comprehensive [Ultralytics ecosystem](https://docs.ultralytics.com/), developers can easily experiment with [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), YOLO11, and YOLO26 to find the perfect performance balance for their specific real-world challenges. diff --git a/docs/en/compare/yolov6-vs-yolox.md b/docs/en/compare/yolov6-vs-yolox.md index 0a9aab72e3e..fe4460a87ab 100644 --- a/docs/en/compare/yolov6-vs-yolox.md +++ b/docs/en/compare/yolov6-vs-yolox.md @@ -6,7 +6,7 @@ keywords: YOLOv6-3.0, YOLOX, object detection, model comparison, computer vision # YOLOv6-3.0 vs YOLOX: Evaluating Industrial Object Detectors -The landscape of computer vision has been heavily shaped by models aiming to bridge the gap between academic research and industrial application. When evaluating [object detection](https://docs.ultralytics.com/tasks/detect/) frameworks tailored for high-performance deployment, **YOLOv6-3.0** and **YOLOX** frequently emerge as prominent contenders. Both models introduce distinct architectural philosophies to maximize throughput and precision, yet they differ significantly in their design choices and primary deployment targets. +The landscape of computer vision has been heavily shaped by models aiming to bridge the gap between academic research and industrial application. When evaluating [object detection](https://docs.ultralytics.com/tasks/detect) frameworks tailored for high-performance deployment, **YOLOv6-3.0** and **YOLOX** frequently emerge as prominent contenders. Both models introduce distinct architectural philosophies to maximize throughput and precision, yet they differ significantly in their design choices and primary deployment targets. @@ -35,7 +35,7 @@ Furthermore, YOLOv6-3.0 utilizes an Anchor-Aided Training (AAT) strategy. This i While YOLOv6 excels on dedicated GPUs, its highly specialized architecture can sometimes result in suboptimal latency when deployed on standard CPUs or low-power edge devices. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## YOLOX: Bridging Research and Industry @@ -51,7 +51,7 @@ Introduced by Megvii, YOLOX represented a significant shift in the YOLO family b YOLOX successfully integrated an anchor-free mechanism with a decoupled head structure. By separating the classification and regression tasks into distinct pathways, YOLOX significantly improved convergence speed and mitigated the conflicting objectives often found in coupled detection heads. -Additionally, YOLOX introduced strong data augmentation strategies (such as MixUp and Mosaic) natively into its training pipeline, drastically improving its robustness when trained from scratch on standard benchmarks like the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). +Additionally, YOLOX introduced strong data augmentation strategies (such as MixUp and Mosaic) natively into its training pipeline, drastically improving its robustness when trained from scratch on standard benchmarks like the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). !!! tip "Decoupled Head Advantage" @@ -101,11 +101,11 @@ YOLOX is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Introducing YOLO26 @@ -115,17 +115,17 @@ Released in January 2026, YOLO26 represents a paradigm shift. It delivers unpara ### Key YOLO26 Innovations -- **End-to-End NMS-Free Design:** Building on concepts pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates the need for Non-Maximum Suppression (NMS) post-processing. This significantly reduces latency variance and simplifies edge deployment. +- **End-to-End NMS-Free Design:** Building on concepts pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates the need for Non-Maximum Suppression (NMS) post-processing. This significantly reduces latency variance and simplifies edge deployment. - **MuSGD Optimizer:** YOLO26 borrows innovations from LLM training stability, utilizing a hybrid MuSGD optimizer (inspired by Moonshot AI's Kimi K2). This enables incredibly stable training dynamics and faster convergence compared to older optimizers. - **Up to 43% Faster CPU Inference:** Unlike YOLOv6, which struggles on non-GPU hardware, YOLO26 is heavily optimized for edge devices. By implementing DFL Removal (Distribution Focal Loss), the output head is simplified, making it incredibly fast on mobile and CPU environments. - **ProgLoss + STAL:** Superior loss functions dramatically improve small object detection, an area where older architectures like YOLOX often struggled. This makes YOLO26 ideal for aerial imagery and IoT sensors. -- **Unmatched Versatility:** While YOLOv6 and YOLOX are strictly detection models, a single YOLO26 architecture natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +- **Unmatched Versatility:** While YOLOv6 and YOLOX are strictly detection models, a single YOLO26 architecture natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } ### Ease of Use and Ecosystem Support -Choosing Ultralytics ensures access to a well-maintained, actively developed ecosystem. The Ultralytics Python package offers a "zero-to-hero" experience, featuring extremely low memory requirements during training compared to bulky transformer models, and seamless exports to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/), [OpenVINO](https://docs.ultralytics.com/integrations/openvino/), and CoreML. +Choosing Ultralytics ensures access to a well-maintained, actively developed ecosystem. The Ultralytics Python package offers a "zero-to-hero" experience, featuring extremely low memory requirements during training compared to bulky transformer models, and seamless exports to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx), [OpenVINO](https://docs.ultralytics.com/integrations/openvino), and CoreML. ```python from ultralytics import YOLO diff --git a/docs/en/compare/yolov7-vs-damo-yolo.md b/docs/en/compare/yolov7-vs-damo-yolo.md index 6ad65a2e850..3b9dd9600c7 100644 --- a/docs/en/compare/yolov7-vs-damo-yolo.md +++ b/docs/en/compare/yolov7-vs-damo-yolo.md @@ -26,9 +26,9 @@ Developed as a continuation of the YOLO family, YOLOv7 introduced the concept of - **Date:** 2022-07-06 - **Arxiv:** [https://arxiv.org/abs/2207.02696](https://arxiv.org/abs/2207.02696) - **GitHub:** [https://github.com/WongKinYiu/yolov7](https://github.com/WongKinYiu/yolov7) -- **Docs:** [https://docs.ultralytics.com/models/yolov7/](https://docs.ultralytics.com/models/yolov7/) +- **Docs:** [https://docs.ultralytics.com/models/yolov7/](https://docs.ultralytics.com/models/yolov7) -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ### DAMO-YOLO Details @@ -76,15 +76,15 @@ The table above demonstrates that YOLOv7 scales well into high-accuracy domains A major distinction between the two architectures lies in their training methodologies. DAMO-YOLO's reliance on distillation means that training a new model from scratch or fine-tuning on a [custom computer vision dataset](https://www.ultralytics.com/blog/custom-training-ultralytics-yolo11-with-computer-vision-datasets) often demands significantly more VRAM and [GPU compute](https://www.ultralytics.com/glossary/gpu-graphics-processing-unit) time. -In contrast, models integrated into the Ultralytics ecosystem, such as YOLOv7 and later versions, are heavily optimized for [memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics/). They allow developers to utilize larger batch sizes on consumer hardware without encountering out-of-memory errors, simplifying the [experiment tracking](https://www.ultralytics.com/glossary/experiment-tracking) and iteration process. +In contrast, models integrated into the Ultralytics ecosystem, such as YOLOv7 and later versions, are heavily optimized for [memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics). They allow developers to utilize larger batch sizes on consumer hardware without encountering out-of-memory errors, simplifying the [experiment tracking](https://www.ultralytics.com/glossary/experiment-tracking) and iteration process. ## The Ultralytics Advantage While both YOLOv7 and DAMO-YOLO offer compelling features, deploying models within the [Ultralytics ecosystem](https://www.ultralytics.com/) provides an unparalleled developer experience. -- **Ease of Use:** The Ultralytics Python package offers a unified, simple API. You can quickly switch between model architectures, start [training loops](https://docs.ultralytics.com/modes/train/), or run [inference](https://docs.ultralytics.com/modes/predict/) with a few lines of code. -- **Well-Maintained Ecosystem:** Ultralytics provides frequent updates, ensuring native compatibility with the latest [PyTorch](https://pytorch.org/) releases and CUDA drivers. It also simplifies exporting models to formats like [ONNX](https://onnx.ai/), [TensorRT](https://developer.nvidia.com/tensorrt), and [OpenVINO](https://docs.ultralytics.com/integrations/openvino/). -- **Versatility:** Unlike DAMO-YOLO, which is strictly an object detector, the Ultralytics ecosystem supports diverse tasks natively. Models from the Ultralytics family can perform standard bounding box detection, [pose estimation](https://docs.ultralytics.com/tasks/pose/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +- **Ease of Use:** The Ultralytics Python package offers a unified, simple API. You can quickly switch between model architectures, start [training loops](https://docs.ultralytics.com/modes/train), or run [inference](https://docs.ultralytics.com/modes/predict) with a few lines of code. +- **Well-Maintained Ecosystem:** Ultralytics provides frequent updates, ensuring native compatibility with the latest [PyTorch](https://pytorch.org/) releases and CUDA drivers. It also simplifies exporting models to formats like [ONNX](https://onnx.ai/), [TensorRT](https://developer.nvidia.com/tensorrt), and [OpenVINO](https://docs.ultralytics.com/integrations/openvino). +- **Versatility:** Unlike DAMO-YOLO, which is strictly an object detector, the Ultralytics ecosystem supports diverse tasks natively. Models from the Ultralytics family can perform standard bounding box detection, [pose estimation](https://docs.ultralytics.com/tasks/pose), [instance segmentation](https://docs.ultralytics.com/tasks/segment), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). ### Code Example: Getting Started Quickly @@ -114,7 +114,7 @@ model.export(format="onnx") While YOLOv7 remains a strong legacy architecture, the field has advanced rapidly. For new deployments, [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) (released January 2026) is the recommended standard, outperforming previous generations in almost every metric. -- **End-to-End NMS-Free Design:** First pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing. This ensures deterministic, ultra-low latency inference critical for robotics and self-driving technologies. +- **End-to-End NMS-Free Design:** First pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing. This ensures deterministic, ultra-low latency inference critical for robotics and self-driving technologies. - **MuSGD Optimizer:** Inspired by advanced LLM training techniques (like Moonshot AI's Kimi K2), this hybrid optimizer blends SGD and Muon to deliver highly stable training and faster convergence across datasets. - **Up to 43% Faster CPU Inference:** By strategically removing Distribution Focal Loss (DFL), YOLO26 significantly boosts performance on edge computing platforms and CPUs. - **ProgLoss + STAL:** These advanced loss functions yield substantial improvements in detecting small objects, making YOLO26 exceptionally well-suited for [aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision) and detailed surveillance. @@ -134,4 +134,4 @@ While YOLOv7 remains a strong legacy architecture, the field has advanced rapidl ### Why Migrate to Modern Ultralytics Models (YOLO11 / YOLO26) -For the vast majority of enterprise applications—from [retail analytics](https://www.ultralytics.com/blog/ai-in-retail-enhancing-customer-experience-using-computer-vision) and [smart manufacturing](https://www.ultralytics.com/blog/improving-manufacturing-with-computer-vision) to healthcare—modern Ultralytics models are unmatched. The integration with the [Ultralytics Platform](https://docs.ultralytics.com/platform/) provides a complete ML pipeline, offering ease of use, superior documentation, robust community support, and multi-task versatility. Whether tracking inventory on a Raspberry Pi or running heavy analytics in the cloud, models like YOLO26 offer the ideal performance balance for the future of computer vision. +For the vast majority of enterprise applications—from [retail analytics](https://www.ultralytics.com/blog/ai-in-retail-enhancing-customer-experience-using-computer-vision) and [smart manufacturing](https://www.ultralytics.com/blog/improving-manufacturing-with-computer-vision) to healthcare—modern Ultralytics models are unmatched. The integration with the [Ultralytics Platform](https://docs.ultralytics.com/platform) provides a complete ML pipeline, offering ease of use, superior documentation, robust community support, and multi-task versatility. Whether tracking inventory on a Raspberry Pi or running heavy analytics in the cloud, models like YOLO26 offer the ideal performance balance for the future of computer vision. diff --git a/docs/en/compare/yolov7-vs-efficientdet.md b/docs/en/compare/yolov7-vs-efficientdet.md index 41fd4ec2988..de75b1c95da 100644 --- a/docs/en/compare/yolov7-vs-efficientdet.md +++ b/docs/en/compare/yolov7-vs-efficientdet.md @@ -23,9 +23,9 @@ Organization: [Institute of Information Science, Academia Sinica, Taiwan](https: Date: 2022-07-06 Arxiv: [YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors](https://arxiv.org/abs/2207.02696) GitHub: [WongKinYiu/yolov7](https://github.com/WongKinYiu/yolov7) -Docs: [Ultralytics YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7/) +Docs: [Ultralytics YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7) -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } **EfficientDet** Authors: Mingxing Tan, Ruoming Pang, and Quoc V. Le @@ -38,7 +38,7 @@ GitHub: [Google AutoML EfficientDet](https://github.com/google/automl/tree/maste ## Architectural Differences and Balanced Analysis -Understanding the fundamental structural differences between these networks is crucial for effective [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/). +Understanding the fundamental structural differences between these networks is crucial for effective [model deployment](https://docs.ultralytics.com/guides/model-deployment-options). ### EfficientDet: Compound Scaling and BiFPN @@ -62,7 +62,7 @@ YOLOv7 prioritized [real-time inference](https://www.ultralytics.com/glossary/re ## Performance Metrics and Benchmarks -The table below contrasts key [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/) enabling developers to assess the trade-offs between speed, parameter count, and accuracy. +The table below contrasts key [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics) enabling developers to assess the trade-offs between speed, parameter count, and accuracy. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | --------------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -78,24 +78,24 @@ The table below contrasts key [performance metrics](https://docs.ultralytics.com | EfficientDet-d6 | 640 | 52.6 | 92.8 | 89.29 | 51.9 | 226.0 | | EfficientDet-d7 | 640 | **53.7** | 122.0 | 128.07 | 51.9 | 325.0 | -As shown, while EfficientDet-d7 achieves a high mAP, its [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) speed severely lags behind YOLOv7 variants, highlighting the latter's dominance in GPU-accelerated [real-time object detection](https://www.ultralytics.com/glossary/real-time-inference). +As shown, while EfficientDet-d7 achieves a high mAP, its [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) speed severely lags behind YOLOv7 variants, highlighting the latter's dominance in GPU-accelerated [real-time object detection](https://www.ultralytics.com/glossary/real-time-inference). ## The Evolution of Object Detection: YOLO26 While YOLOv7 and EfficientDet laid vital groundwork, the landscape of [vision AI](https://www.ultralytics.com/blog/a-quick-overview-of-vision-ai-and-how-it-works) evolves rapidly. For modern applications requiring the absolute pinnacle of efficiency and accuracy, we highly recommend upgrading to **YOLO26**, released in January 2026. -YOLO26 addresses the inherent limitations of previous generations, offering unprecedented [versatility](https://docs.ultralytics.com/tasks/) across [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/). +YOLO26 addresses the inherent limitations of previous generations, offering unprecedented [versatility](https://docs.ultralytics.com/tasks) across [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose). [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } ### Key YOLO26 Innovations -- **End-to-End NMS-Free Design:** YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing. Pioneered initially in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), this simplifies deployment logic and guarantees consistent, low-latency execution regardless of object density. +- **End-to-End NMS-Free Design:** YOLO26 natively eliminates Non-Maximum Suppression (NMS) post-processing. Pioneered initially in [YOLOv10](https://docs.ultralytics.com/models/yolov10), this simplifies deployment logic and guarantees consistent, low-latency execution regardless of object density. - **DFL Removal:** By removing the Distribution Focal Loss (DFL), the model architecture is vastly simplified, enhancing compatibility with highly constrained [edge computing](https://www.ultralytics.com/glossary/edge-computing) environments. - **Up to 43% Faster CPU Inference:** Heavily optimized for environments lacking dedicated GPUs, making it exponentially faster than EfficientDet on lightweight hardware. -- **MuSGD Optimizer:** Inspired by large language model techniques (such as Moonshot AI's Kimi K2), this hybrid of SGD and Muon brings LLM-level stability and rapid convergence to [computer vision training](https://docs.ultralytics.com/modes/train/). +- **MuSGD Optimizer:** Inspired by large language model techniques (such as Moonshot AI's Kimi K2), this hybrid of SGD and Muon brings LLM-level stability and rapid convergence to [computer vision training](https://docs.ultralytics.com/modes/train). - **ProgLoss + STAL:** These advanced loss functions deliver remarkable improvements in small-object recognition, a critical feature for [aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision) and [drone applications](https://www.ultralytics.com/blog/build-ai-powered-drone-applications-with-ultralytics-yolo11). -- **Task-Specific Improvements:** Includes Semantic segmentation loss and multi-scale proto for segmentation tasks, Residual Log-Likelihood Estimation (RLE) for complex Pose estimation, and a specialized angle loss tailored to fix [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) boundary issues. +- **Task-Specific Improvements:** Includes Semantic segmentation loss and multi-scale proto for segmentation tasks, Residual Log-Likelihood Estimation (RLE) for complex Pose estimation, and a specialized angle loss tailored to fix [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) boundary issues. For teams currently using legacy systems, transitioning to the [Ultralytics Platform](https://platform.ultralytics.com/) unlocks a streamlined workflow where these cutting-edge models can be trained and deployed with ease. Developers may also explore previous robust iterations like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) depending on specific backward-compatibility requirements. @@ -103,11 +103,11 @@ For teams currently using legacy systems, transitioning to the [Ultralytics Plat One of the defining characteristics of Ultralytics models is the sheer **Ease of Use**. Unlike the complex, multi-dependency setup required for EfficientDet's TensorFlow AutoML environments, Ultralytics provides a simple, Pythonic API. -This environment minimizes [CUDA memory usage](https://docs.ultralytics.com/guides/yolo-performance-metrics/) during training, ensuring that even large datasets can be processed efficiently without Out-Of-Memory (OOM) errors commonly seen in bulky [Transformer-based](https://www.ultralytics.com/glossary/transformer) architectures. +This environment minimizes [CUDA memory usage](https://docs.ultralytics.com/guides/yolo-performance-metrics) during training, ensuring that even large datasets can be processed efficiently without Out-Of-Memory (OOM) errors commonly seen in bulky [Transformer-based](https://www.ultralytics.com/glossary/transformer) architectures. ### Code Example: Getting Started with Ultralytics -The following snippet demonstrates how developers can leverage the [Ultralytics package](https://docs.ultralytics.com/usage/python/) to train a state-of-the-art YOLO26 model seamlessly out of the box. +The following snippet demonstrates how developers can leverage the [Ultralytics package](https://docs.ultralytics.com/usage/python) to train a state-of-the-art YOLO26 model seamlessly out of the box. ```python from ultralytics import YOLO @@ -134,7 +134,7 @@ model.export(format="openvino") !!! tip "Exporting for Production" - Models trained via the Ultralytics API can be instantly exported to various production formats like [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) or [ONNX](https://docs.ultralytics.com/integrations/onnx/), ensuring high throughput regardless of your target hardware. + Models trained via the Ultralytics API can be instantly exported to various production formats like [OpenVINO](https://docs.ultralytics.com/integrations/openvino) or [ONNX](https://docs.ultralytics.com/integrations/onnx), ensuring high throughput regardless of your target hardware. ## Ideal Use Cases and Real-World Applications diff --git a/docs/en/compare/yolov7-vs-pp-yoloe.md b/docs/en/compare/yolov7-vs-pp-yoloe.md index 3a832afa07d..99ad3c2fd4b 100644 --- a/docs/en/compare/yolov7-vs-pp-yoloe.md +++ b/docs/en/compare/yolov7-vs-pp-yoloe.md @@ -31,7 +31,7 @@ Organization: Institute of Information Science, Academia Sinica, Taiwan Date: 2022-07-06 Arxiv: [https://arxiv.org/abs/2207.02696](https://arxiv.org/abs/2207.02696) -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ### PP-YOLOE+ Architecture Highlights @@ -72,7 +72,7 @@ Choosing the right model often comes down to the specific constraints of your ha !!! tip "Optimizing for Production" - When deploying these models, leveraging export formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) or [ONNX](https://docs.ultralytics.com/integrations/onnx/) can significantly reduce latency compared to native PyTorch inference. + When deploying these models, leveraging export formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) or [ONNX](https://docs.ultralytics.com/integrations/onnx) can significantly reduce latency compared to native PyTorch inference. ## The Ultralytics Advantage @@ -97,11 +97,11 @@ model.export(format="engine") # TensorRT export ### Resource Efficiency -A major strength of Ultralytics YOLO models is their lower **memory requirements** during both training and inference. This efficiency allows researchers and developers to use larger batch sizes on consumer-grade hardware, accelerating the training process compared to heavier models or complex Transformer architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +A major strength of Ultralytics YOLO models is their lower **memory requirements** during both training and inference. This efficiency allows researchers and developers to use larger batch sizes on consumer-grade hardware, accelerating the training process compared to heavier models or complex Transformer architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). ### Ecosystem and Versatility -The Ultralytics ecosystem is exceptionally **well-maintained**, featuring frequent updates, extensive documentation, and native support for diverse tasks beyond standard detection. With Ultralytics, a single framework supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), providing unmatched **versatility** that competing models often lack. +The Ultralytics ecosystem is exceptionally **well-maintained**, featuring frequent updates, extensive documentation, and native support for diverse tasks beyond standard detection. With Ultralytics, a single framework supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb), providing unmatched **versatility** that competing models often lack. ## The Future of Vision AI: YOLO26 @@ -109,7 +109,7 @@ As computer vision rapidly evolves, newer architectures have emerged that redefi **Key YOLO26 Innovations:** -- **End-to-End NMS-Free Design:** YOLO26 eliminates Non-Maximum Suppression (NMS) post-processing. This natively end-to-end approach drastically simplifies deployment logic and reduces variable latency, a breakthrough first introduced in [YOLOv10](https://docs.ultralytics.com/models/yolov10/). +- **End-to-End NMS-Free Design:** YOLO26 eliminates Non-Maximum Suppression (NMS) post-processing. This natively end-to-end approach drastically simplifies deployment logic and reduces variable latency, a breakthrough first introduced in [YOLOv10](https://docs.ultralytics.com/models/yolov10). - **Unprecedented Edge Performance:** By removing Distribution Focal Loss (DFL), YOLO26 achieves up to **43% faster CPU inference**, making it superior for IoT and edge devices compared to previous generations. - **Advanced Training Dynamics:** The integration of the **MuSGD Optimizer**—inspired by LLM innovations like Moonshot AI's Kimi K2—ensures more stable training and faster convergence. - **Superior Small Object Detection:** Enhanced loss functions, specifically **ProgLoss + STAL**, address historical weaknesses in recognizing small objects, crucial for applications like [aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision). @@ -129,4 +129,4 @@ Choosing between these architectures often depends on the specific deployment en - **Robotics Integration:** Ideal for [integrating computer vision in robotics](https://www.ultralytics.com/blog/integrating-computer-vision-in-robotics-with-ultalytics-yolo11), allowing for fast decision-making in dynamic environments. - **Academic Research:** Widely supported and frequently used as a reliable baseline in PyTorch-based research. -While older models hold historical significance, transitioning to modern architectures like **YOLO26** or [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) via the [Ultralytics Platform](https://docs.ultralytics.com/platform/) ensures access to the latest optimizations, the simplest training workflows, and the broadest multi-task support available today. +While older models hold historical significance, transitioning to modern architectures like **YOLO26** or [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) via the [Ultralytics Platform](https://docs.ultralytics.com/platform) ensures access to the latest optimizations, the simplest training workflows, and the broadest multi-task support available today. diff --git a/docs/en/compare/yolov7-vs-rtdetr.md b/docs/en/compare/yolov7-vs-rtdetr.md index 57a1b9e368e..7da2c969bf0 100644 --- a/docs/en/compare/yolov7-vs-rtdetr.md +++ b/docs/en/compare/yolov7-vs-rtdetr.md @@ -28,15 +28,15 @@ GitHub: [WongKinYiu/yolov7](https://github.com/WongKinYiu/yolov7) ### Architecture and Strengths -YOLOv7 thrives on its Extended Efficient Layer Aggregation Network (E-ELAN) architecture. This structural design enables the model to learn more diverse features without destroying the original gradient path. Furthermore, it incorporates planned re-parameterized convolutions, which optimize inference speed without degrading accuracy. Its decoupled head structure allows it to achieve impressive trade-offs between speed and accuracy, making it highly suitable for [real-time object detection](https://docs.ultralytics.com/tasks/detect/) tasks on server-grade GPUs. +YOLOv7 thrives on its Extended Efficient Layer Aggregation Network (E-ELAN) architecture. This structural design enables the model to learn more diverse features without destroying the original gradient path. Furthermore, it incorporates planned re-parameterized convolutions, which optimize inference speed without degrading accuracy. Its decoupled head structure allows it to achieve impressive trade-offs between speed and accuracy, making it highly suitable for [real-time object detection](https://docs.ultralytics.com/tasks/detect) tasks on server-grade GPUs. -YOLOv7 is also highly versatile. Beyond standard bounding box detection, the repository offers branches for [pose estimation](https://docs.ultralytics.com/tasks/pose/) and [instance segmentation](https://docs.ultralytics.com/tasks/segment/), demonstrating its adaptability. +YOLOv7 is also highly versatile. Beyond standard bounding box detection, the repository offers branches for [pose estimation](https://docs.ultralytics.com/tasks/pose) and [instance segmentation](https://docs.ultralytics.com/tasks/segment), demonstrating its adaptability. ### Limitations Like many legacy CNN models, YOLOv7 relies on Non-Maximum Suppression (NMS) for post-processing. NMS introduces variable latency, especially in crowded scenes, which can complicate strict real-time guarantees on edge devices. -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ## RTDETRv2: Advancing Real-Time Transformers @@ -55,7 +55,7 @@ RTDETRv2 represents a significant step forward for Vision Transformers. It lever ### Limitations -Despite its advancements, RTDETRv2 carries the traditional burdens of transformer-based architectures. It demands significantly higher CUDA memory during both training and inference compared to CNNs. Additionally, its training convergence times are noticeably longer, requiring vast amounts of high-quality annotated data (like the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/)) and heavy computational resources. +Despite its advancements, RTDETRv2 carries the traditional burdens of transformer-based architectures. It demands significantly higher CUDA memory during both training and inference compared to CNNs. Additionally, its training convergence times are noticeably longer, requiring vast amounts of high-quality annotated data (like the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco)) and heavy computational resources. [Learn more about RTDETRv2](https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetrv2_pytorch#readme){ .md-button } @@ -83,11 +83,11 @@ While YOLOv7 and RTDETRv2 offer robust capabilities, deploying them in productio ### Unmatched Versatility and Memory Efficiency -Unlike rigid transformer models that consume massive amounts of VRAM, Ultralytics YOLO models maintain strict memory efficiency. This enables rapid [model training](https://docs.ultralytics.com/modes/train/) on accessible hardware. The ecosystem inherently supports multiple computer vision tasks from a single codebase, including [image classification](https://docs.ultralytics.com/tasks/classify/) and [oriented bounding box (OBB) detection](https://docs.ultralytics.com/tasks/obb/), offering a flexibility that RTDETRv2 currently lacks. +Unlike rigid transformer models that consume massive amounts of VRAM, Ultralytics YOLO models maintain strict memory efficiency. This enables rapid [model training](https://docs.ultralytics.com/modes/train) on accessible hardware. The ecosystem inherently supports multiple computer vision tasks from a single codebase, including [image classification](https://docs.ultralytics.com/tasks/classify) and [oriented bounding box (OBB) detection](https://docs.ultralytics.com/tasks/obb), offering a flexibility that RTDETRv2 currently lacks. ### Seamless Deployment -Moving from research to production requires robust deployment options. The Ultralytics API natively handles one-click [model export](https://docs.ultralytics.com/modes/export/) to industry-standard formats. Whether you are targeting [ONNX](https://docs.ultralytics.com/integrations/onnx/) for cross-platform compatibility or [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) for maxed-out GPU acceleration, the pipeline is fully automated and reliable. +Moving from research to production requires robust deployment options. The Ultralytics API natively handles one-click [model export](https://docs.ultralytics.com/modes/export) to industry-standard formats. Whether you are targeting [ONNX](https://docs.ultralytics.com/integrations/onnx) for cross-platform compatibility or [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) for maxed-out GPU acceleration, the pipeline is fully automated and reliable. ## The Ultimate Upgrade: Ultralytics YOLO26 @@ -97,7 +97,7 @@ For developers debating between YOLOv7 and RTDETRv2, the optimal path forward is YOLO26 introduces groundbreaking innovations tailored for both server and edge deployments: -- **End-to-End NMS-Free Design:** First pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates NMS post-processing. This ensures the deterministic latency of RTDETRv2 without the burdensome computational overhead of a transformer. +- **End-to-End NMS-Free Design:** First pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates NMS post-processing. This ensures the deterministic latency of RTDETRv2 without the burdensome computational overhead of a transformer. - **MuSGD Optimizer:** Inspired by large language model training techniques (such as Moonshot AI's Kimi K2), YOLO26 utilizes a hybrid of SGD and Muon. This delivers unprecedented training stability and significantly faster convergence times compared to standard AdamW implementations used by ViTs. - **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, directly competing with the multi-scale feature advantages of RTDETRv2, which is critical for [robotic automation](https://www.ultralytics.com/blog/from-algorithms-to-automation-ais-role-in-robotics). - **Edge Optimization & DFL Removal:** By removing Distribution Focal Loss (DFL), YOLO26 streamlines the output head, leading to up to **43% faster CPU inference**—making it infinitely more deployable on edge devices than heavy transformer models. @@ -140,4 +140,4 @@ Choosing the right architecture depends heavily on deployment constraints and ha !!! tip "Explore More Models" - Interested in how other architectures stack up? Explore our deep dives into previous generations like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), or learn how to leverage [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/) to maximize your project's accuracy. + Interested in how other architectures stack up? Explore our deep dives into previous generations like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), or learn how to leverage [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning) to maximize your project's accuracy. diff --git a/docs/en/compare/yolov7-vs-yolo11.md b/docs/en/compare/yolov7-vs-yolo11.md index c0da8e57537..7b176f7d7b2 100644 --- a/docs/en/compare/yolov7-vs-yolo11.md +++ b/docs/en/compare/yolov7-vs-yolo11.md @@ -6,7 +6,7 @@ keywords: YOLO11, YOLOv7, object detection, model comparison, YOLO models, deep # YOLOv7 vs YOLO11: A Comprehensive Technical Comparison -The landscape of computer vision has rapidly evolved over the past few years. For developers and researchers choosing the right object detection framework, understanding the architectural and practical differences between generation-defining models is critical. This guide provides a detailed technical comparison between the academic breakthrough of [YOLOv7](https://docs.ultralytics.com/models/yolov7/) and the highly refined, production-ready [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11). +The landscape of computer vision has rapidly evolved over the past few years. For developers and researchers choosing the right object detection framework, understanding the architectural and practical differences between generation-defining models is critical. This guide provides a detailed technical comparison between the academic breakthrough of [YOLOv7](https://docs.ultralytics.com/models/yolov7) and the highly refined, production-ready [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11). @@ -17,7 +17,7 @@ The landscape of computer vision has rapidly evolved over the past few years. Fo **YOLOv7**, released on July 6, 2022, by authors Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao from the [Institute of Information Science at Academia Sinica](https://www.iis.sinica.edu.tw/en/index.html), introduced several novel concepts to the field. Detailed in their [YOLOv7 research paper published on arXiv](https://arxiv.org/abs/2207.02696), the model focuses heavily on a "trainable bag-of-freebies" approach and Extended Efficient Layer Aggregation Networks (E-ELAN). These architectural choices were specifically designed to maximize gradient path efficiency, making it a powerful tool for academic benchmarking on high-end GPUs. -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } **YOLO11**, developed by Glenn Jocher and Jing Qiu at [Ultralytics](https://www.ultralytics.com/about), was released on September 27, 2024. YOLO11 shifts the focus from pure architectural complexity to a holistic, developer-first ecosystem. Hosted on the [Ultralytics GitHub repository](https://github.com/ultralytics/ultralytics), YOLO11 features an optimized anchor-free design that drastically reduces memory consumption during both training and inference. It is natively integrated into the [Ultralytics Platform](https://platform.ultralytics.com), offering unparalleled ease of use from dataset annotation to edge deployment. @@ -58,7 +58,7 @@ Using the original [YOLOv7 open-source codebase](https://github.com/WongKinYiu/y ### Training YOLO11 -YOLO11 is deeply integrated into the `ultralytics` Python package, simplifying the machine learning lifecycle. Training an [object detection model](https://docs.ultralytics.com/tasks/detect/) takes only a few lines of code, and the framework natively handles data downloading, hyperparameter tuning, and caching. +YOLO11 is deeply integrated into the `ultralytics` Python package, simplifying the machine learning lifecycle. Training an [object detection model](https://docs.ultralytics.com/tasks/detect) takes only a few lines of code, and the framework natively handles data downloading, hyperparameter tuning, and caching. ```python from ultralytics import YOLO @@ -73,7 +73,7 @@ results = model.train(data="coco8.yaml", epochs=50, imgsz=640) export_path = model.export(format="onnx") ``` -Furthermore, YOLO11 boasts extreme versatility. By simply changing the model suffix, developers can instantly transition from detection to [instance segmentation mapping](https://docs.ultralytics.com/tasks/segment/), [pose estimation tracking](https://docs.ultralytics.com/tasks/pose/), or [Oriented Bounding Box (OBB) recognition](https://docs.ultralytics.com/tasks/obb/)—a level of native multi-task support that YOLOv7 lacks. +Furthermore, YOLO11 boasts extreme versatility. By simply changing the model suffix, developers can instantly transition from detection to [instance segmentation mapping](https://docs.ultralytics.com/tasks/segment), [pose estimation tracking](https://docs.ultralytics.com/tasks/pose), or [Oriented Bounding Box (OBB) recognition](https://docs.ultralytics.com/tasks/obb)—a level of native multi-task support that YOLOv7 lacks. !!! info "Simplified Exports" @@ -83,9 +83,9 @@ Furthermore, YOLO11 boasts extreme versatility. By simply changing the model suf Understanding the strengths of each model helps dictate their best use cases. -- [Legacy Benchmark Reproduction](https://docs.ultralytics.com/compare/yolov7-vs-yolov8/): **YOLOv7** remains useful for academic researchers who need to reproduce specific 2022 benchmarks or study the effects of re-parameterization techniques on anchor-based networks. +- [Legacy Benchmark Reproduction](https://docs.ultralytics.com/compare/yolov7-vs-yolov8): **YOLOv7** remains useful for academic researchers who need to reproduce specific 2022 benchmarks or study the effects of re-parameterization techniques on anchor-based networks. - [Commercial Production Environments](https://www.ultralytics.com/solutions/ai-in-retail): **YOLO11** is the clear choice for enterprise systems. Its stability, active maintenance, and integration with the cloud-based [Ultralytics Platform interface](https://platform.ultralytics.com) make it ideal for managing large-scale retail analytics, security monitoring, and manufacturing quality control. -- [Resource-Constrained Edge Computing](https://docs.ultralytics.com/guides/raspberry-pi/): The incredibly lightweight YOLO11n variant is specifically designed for low-power edge devices, running efficiently on a [Raspberry Pi system](https://www.raspberrypi.org/) or [NVIDIA Jetson modules](https://developer.nvidia.com/embedded-computing). +- [Resource-Constrained Edge Computing](https://docs.ultralytics.com/guides/raspberry-pi): The incredibly lightweight YOLO11n variant is specifically designed for low-power edge devices, running efficiently on a [Raspberry Pi system](https://www.raspberrypi.org/) or [NVIDIA Jetson modules](https://developer.nvidia.com/embedded-computing). ## Looking Forward: The Paradigm Shift of YOLO26 @@ -96,8 +96,8 @@ Released in January 2026, YOLO26 introduces several groundbreaking features that - **Natively NMS-Free Architecture:** YOLO26 eliminates the need for Non-Maximum Suppression post-processing. This end-to-end design simplifies deployment pipelines and dramatically reduces latency variability. - **Up to 43% Faster CPU Inference:** By strategically removing the Distribution Focal Loss (DFL) module, YOLO26 is heavily optimized for edge devices and environments without dedicated GPUs. - **MuSGD Optimizer Integration:** Inspired by advanced LLM training techniques from [Moonshot AI](https://www.moonshot.cn/), this hybrid optimizer ensures unprecedented training stability and faster convergence rates. -- **Superior Small Object Detection:** The introduction of ProgLoss and STAL loss functions provides critical accuracy boosts for identifying minute details, perfect for analyzing [drone aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and complex IoT sensor data. +- **Superior Small Object Detection:** The introduction of ProgLoss and STAL loss functions provides critical accuracy boosts for identifying minute details, perfect for analyzing [drone aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and complex IoT sensor data. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } -For users interested in transformer-based architectures or alternative paradigms, the Ultralytics documentation also covers models like the [RT-DETR transformer detector](https://docs.ultralytics.com/models/rtdetr/) and the [YOLO-World open-vocabulary model](https://docs.ultralytics.com/models/yolo-world/). +For users interested in transformer-based architectures or alternative paradigms, the Ultralytics documentation also covers models like the [RT-DETR transformer detector](https://docs.ultralytics.com/models/rtdetr) and the [YOLO-World open-vocabulary model](https://docs.ultralytics.com/models/yolo-world). diff --git a/docs/en/compare/yolov7-vs-yolo26.md b/docs/en/compare/yolov7-vs-yolo26.md index 30cb4834919..0d8ccab876c 100644 --- a/docs/en/compare/yolov7-vs-yolo26.md +++ b/docs/en/compare/yolov7-vs-yolo26.md @@ -24,11 +24,11 @@ Introduced in mid-2022, YOLOv7 pushed the boundaries of what was possible on GPU - **Date:** 2022-07-06 - **Arxiv:** [2207.02696](https://arxiv.org/abs/2207.02696) - **GitHub:** [WongKinYiu/yolov7](https://github.com/WongKinYiu/yolov7) -- **Docs:** [Ultralytics YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7/) +- **Docs:** [Ultralytics YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7) YOLOv7 introduced the concept of trainable "bag-of-freebies," which heavily utilized re-parameterization techniques and extended efficient layer aggregation networks (E-ELAN). This allowed the model to learn more diverse features and continuously improve the learning capability of the network without destroying the original gradient path. While it achieved an impressive state-of-the-art benchmark on COCO at the time, its architecture remains heavily reliant on anchor-based outputs and requires complex [Non-Maximum Suppression (NMS)](https://en.wikipedia.org/wiki/NMS) post-processing, which can introduce latency bottlenecks during deployment. -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ## YOLO26: The Edge-First Vision AI Standard @@ -44,12 +44,12 @@ Released in January 2026, Ultralytics YOLO26 represents a paradigm shift, entire YOLO26 is built from the ground up to solve modern engineering challenges. Its architecture brings several critical innovations that significantly outpace its predecessors: -- **End-to-End NMS-Free Design:** YOLO26 eliminates NMS post-processing natively, a breakthrough approach first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/). This results in a faster, much simpler deployment pipeline, avoiding the variable latency typically caused by crowded scenes. +- **End-to-End NMS-Free Design:** YOLO26 eliminates NMS post-processing natively, a breakthrough approach first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10). This results in a faster, much simpler deployment pipeline, avoiding the variable latency typically caused by crowded scenes. - **DFL Removal:** By removing the Distribution Focal Loss (DFL), the model is radically simplified for export, offering vastly better compatibility with edge devices and low-power IoT hardware. - **Up to 43% Faster CPU Inference:** Thanks to the architectural simplifications and structural pruning, YOLO26 is specifically optimized for edge computing and devices without dedicated GPUs, easily outperforming older architectures on standard processors. - **MuSGD Optimizer:** Inspired by large language model training techniques (specifically Moonshot AI's Kimi K2), YOLO26 uses the MuSGD optimizer—a hybrid of [Stochastic Gradient Descent](https://en.wikipedia.org/wiki/Stochastic_gradient_descent) and Muon. This brings unparalleled training stability and much faster convergence to computer vision tasks. -- **ProgLoss + STAL:** The introduction of these advanced loss functions yields notable improvements in small-object recognition, which is critical for [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/), robotics, and automated quality inspection. -- **Task-Specific Improvements:** Beyond standard [object detection](https://docs.ultralytics.com/tasks/detect/), YOLO26 introduces multi-scale proto and specialized semantic segmentation loss for [segmentation tasks](https://docs.ultralytics.com/tasks/segment/), Residual Log-Likelihood Estimation (RLE) for [pose estimation](https://docs.ultralytics.com/tasks/pose/), and specialized angle loss algorithms to resolve boundary issues in [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +- **ProgLoss + STAL:** The introduction of these advanced loss functions yields notable improvements in small-object recognition, which is critical for [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone), robotics, and automated quality inspection. +- **Task-Specific Improvements:** Beyond standard [object detection](https://docs.ultralytics.com/tasks/detect), YOLO26 introduces multi-scale proto and specialized semantic segmentation loss for [segmentation tasks](https://docs.ultralytics.com/tasks/segment), Residual Log-Likelihood Estimation (RLE) for [pose estimation](https://docs.ultralytics.com/tasks/pose), and specialized angle loss algorithms to resolve boundary issues in [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -121,6 +121,6 @@ Choosing the right architecture depends entirely on your production constraints. YOLOv7 remains a valuable tool for academic benchmarking against 2022 standards. If your infrastructure utilizes deep legacy [CUDA pipelines](https://developer.nvidia.com/cuda) heavily hardcoded to YOLOv7's specific anchor outputs and you cannot allocate resources for refactoring, it will continue to function as a robust baseline detector. **When to choose YOLO26:** -For any new project, YOLO26 is the definitive choice. Its NMS-free architecture makes it perfect for low-latency [autonomous navigation](https://docs.ultralytics.com/datasets/detect/argoverse/) and real-time security systems. The removal of DFL and massive CPU speed boosts make it the undisputed champion for edge AI deployments, such as deploying on a [Raspberry Pi](https://www.raspberrypi.org/) or inside consumer electronics. Furthermore, the ProgLoss + STAL enhancements make it highly adept at detecting tiny anomalies in manufacturing quality assurance or satellite imaging. +For any new project, YOLO26 is the definitive choice. Its NMS-free architecture makes it perfect for low-latency [autonomous navigation](https://docs.ultralytics.com/datasets/detect/argoverse) and real-time security systems. The removal of DFL and massive CPU speed boosts make it the undisputed champion for edge AI deployments, such as deploying on a [Raspberry Pi](https://www.raspberrypi.org/) or inside consumer electronics. Furthermore, the ProgLoss + STAL enhancements make it highly adept at detecting tiny anomalies in manufacturing quality assurance or satellite imaging. Ultimately, YOLO26 provides developers with an unmatched blend of accuracy, speed, and simplicity, backed by the comprehensive support of the open-source community. diff --git a/docs/en/compare/yolov7-vs-yolov10.md b/docs/en/compare/yolov7-vs-yolov10.md index 47f1e30f060..855e3b25930 100644 --- a/docs/en/compare/yolov7-vs-yolov10.md +++ b/docs/en/compare/yolov7-vs-yolov10.md @@ -23,7 +23,7 @@ Released on July 6, 2022, by researchers Chien-Yao Wang, Alexey Bochkovskiy, and YOLOv7 leverages an extended efficient layer aggregation network (E-ELAN) to guide the network in learning diverse features without destroying the original gradient path. This makes it a robust choice for academic research benchmarks and systems heavily reliant on standard high-end GPUs. -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ### YOLOv10: Real-Time End-to-End Detection @@ -31,11 +31,11 @@ Developed by Ao Wang and his team at [Tsinghua University](https://www.tsinghua. YOLOv10 introduced consistent dual assignments for NMS-free training, fundamentally altering the post-processing pipeline. By deploying a holistic efficiency-accuracy driven model design strategy, YOLOv10 reduces computational redundancy. This results in an architecture uniquely tailored for edge devices requiring extremely low latency. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } !!! tip "NMS-Free Architecture" - The removal of Non-Maximum Suppression (NMS) in YOLOv10 allows the entire model to be exported as a single computational graph. This vastly simplifies deployment using runtimes like [TensorRT](https://developer.nvidia.com/tensorrt) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino/). + The removal of Non-Maximum Suppression (NMS) in YOLOv10 allows the entire model to be exported as a single computational graph. This vastly simplifies deployment using runtimes like [TensorRT](https://developer.nvidia.com/tensorrt) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino). ## Performance and Metrics Comparison @@ -74,9 +74,9 @@ The NMS-free, lightweight design of YOLOv10 shines in constrained environments. While both models have strong academic roots, their true potential is unlocked when utilized within the unified [Ultralytics Platform](https://platform.ultralytics.com). Developing computer vision models from scratch is notoriously difficult, but the Ultralytics ecosystem provides an unparalleled experience for machine learning engineers. - **Ease of Use:** The Ultralytics Python API provides a unified interface. You can train, validate, and export models with just a few lines of code, avoiding the complex dependency nightmares associated with typical academic repositories. -- **Well-Maintained Ecosystem:** Ultralytics guarantees that the underlying code is actively developed. Users benefit from seamless integrations with popular ML tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) for logging, or [Hugging Face](https://docs.ultralytics.com/integrations/gradio/) for fast web demos. +- **Well-Maintained Ecosystem:** Ultralytics guarantees that the underlying code is actively developed. Users benefit from seamless integrations with popular ML tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) for logging, or [Hugging Face](https://docs.ultralytics.com/integrations/gradio) for fast web demos. - **Memory Requirements:** Transformer-based object detectors often consume massive amounts of CUDA memory during training. In contrast, Ultralytics YOLO models require far less memory, allowing for much larger [batch sizes](https://www.ultralytics.com/glossary/batch-size) on consumer-grade hardware. -- **Versatility:** The Ultralytics pipeline is not restricted to standard bounding boxes. It seamlessly supports [pose estimation](https://docs.ultralytics.com/tasks/pose/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), and oriented bounding boxes across supported model families like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8). +- **Versatility:** The Ultralytics pipeline is not restricted to standard bounding boxes. It seamlessly supports [pose estimation](https://docs.ultralytics.com/tasks/pose), [instance segmentation](https://docs.ultralytics.com/tasks/segment), and oriented bounding boxes across supported model families like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8). ### Streamlined Training Example @@ -120,11 +120,11 @@ YOLOv10 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Future: Introducing YOLO26 diff --git a/docs/en/compare/yolov7-vs-yolov5.md b/docs/en/compare/yolov7-vs-yolov5.md index 4301745a518..715dc5acf9f 100644 --- a/docs/en/compare/yolov7-vs-yolov5.md +++ b/docs/en/compare/yolov7-vs-yolov5.md @@ -29,9 +29,9 @@ You can explore the source code on the [YOLOv5 GitHub repository](https://github ### YOLOv7 Introduced by Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao from the Institute of Information Science, Academia Sinica, Taiwan on July 6, 2022. YOLOv7 focused heavily on architectural innovations like Extended Efficient Layer Aggregation Networks (E-ELAN) and a trainable "bag-of-freebies" to push the state-of-the-art in accuracy. -Details can be found in their [official Arxiv paper](https://arxiv.org/abs/2207.02696) and the [YOLOv7 GitHub repository](https://github.com/WongKinYiu/yolov7). For seamless integration, check out the [Ultralytics YOLOv7 documentation](https://docs.ultralytics.com/models/yolov7/). +Details can be found in their [official Arxiv paper](https://arxiv.org/abs/2207.02696) and the [YOLOv7 GitHub repository](https://github.com/WongKinYiu/yolov7). For seamless integration, check out the [Ultralytics YOLOv7 documentation](https://docs.ultralytics.com/models/yolov7). -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } !!! tip "Seamless Experimentation" @@ -41,7 +41,7 @@ Details can be found in their [official Arxiv paper](https://arxiv.org/abs/2207. ### Ultralytics YOLOv5 Design -YOLOv5 utilizes a modified CSPDarknet53 backbone paired with a Path Aggregation Network (PANet) neck. This design is highly optimized for rapid [feature extraction](https://www.ultralytics.com/glossary/feature-extraction) and memory efficiency. Unlike older architectures or heavy transformer models, YOLOv5 requires significantly less CUDA memory during training, allowing for larger [batch sizes](https://www.ultralytics.com/glossary/batch-size) on standard consumer-grade GPUs. Furthermore, the Ultralytics framework inherently supports a wide variety of tasks beyond standard bounding boxes, including [image segmentation](https://docs.ultralytics.com/tasks/segment/) and [image classification](https://docs.ultralytics.com/tasks/classify/). +YOLOv5 utilizes a modified CSPDarknet53 backbone paired with a Path Aggregation Network (PANet) neck. This design is highly optimized for rapid [feature extraction](https://www.ultralytics.com/glossary/feature-extraction) and memory efficiency. Unlike older architectures or heavy transformer models, YOLOv5 requires significantly less CUDA memory during training, allowing for larger [batch sizes](https://www.ultralytics.com/glossary/batch-size) on standard consumer-grade GPUs. Furthermore, the Ultralytics framework inherently supports a wide variety of tasks beyond standard bounding boxes, including [image segmentation](https://docs.ultralytics.com/tasks/segment) and [image classification](https://docs.ultralytics.com/tasks/classify). ### YOLOv7 Design @@ -49,7 +49,7 @@ YOLOv7 introduced several structural re-parameterizations and the E-ELAN archite ## Performance Analysis -When comparing these models, developers must balance mAPval, inference speed, and computational complexity (FLOPs). The table below demonstrates the performance of both architectures evaluated on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). +When comparing these models, developers must balance mAPval, inference speed, and computational complexity (FLOPs). The table below demonstrates the performance of both architectures evaluated on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -73,8 +73,8 @@ When comparing these models, developers must balance mAPval, inferenc A model's architecture is only half the equation; the ecosystem surrounding it dictates its real-world viability. This is where Ultralytics models truly shine. **Ease of Use:** Ultralytics provides a unified, highly intuitive Python API. You can train, validate, and deploy models with minimal boilerplate, backed by extensive [official documentation](https://docs.ultralytics.com/). -**Well-Maintained Ecosystem:** Active development ensures constant updates, bug fixes, and seamless integration with modern tracking tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/). -**Training Efficiency:** Utilizing optimized data loaders and [smart caching](https://docs.ultralytics.com/guides/preprocessing_annotated_data/), YOLOv5 drastically reduces training times. Moreover, ready-to-use pre-trained weights accelerate transfer learning across various domains. +**Well-Maintained Ecosystem:** Active development ensures constant updates, bug fixes, and seamless integration with modern tracking tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases). +**Training Efficiency:** Utilizing optimized data loaders and [smart caching](https://docs.ultralytics.com/guides/preprocessing_annotated_data), YOLOv5 drastically reduces training times. Moreover, ready-to-use pre-trained weights accelerate transfer learning across various domains. ### Code Example: Streamlined Training @@ -102,9 +102,9 @@ success = model.export(format="onnx") ### When to Choose YOLOv5 -- **Production Deployments:** Ideal for commercial applications requiring high stability, straightforward [model deployment options](https://docs.ultralytics.com/guides/model-deployment-options/), and wide cross-platform compatibility. +- **Production Deployments:** Ideal for commercial applications requiring high stability, straightforward [model deployment options](https://docs.ultralytics.com/guides/model-deployment-options), and wide cross-platform compatibility. - **Edge Devices:** The smaller variants (YOLOv5n and YOLOv5s) run exceptionally well on mobile phones and embedded systems. -- **Multi-Task Requirements:** If your project needs to evolve from simple detection to [pose estimation](https://docs.ultralytics.com/tasks/pose/) or segmentation using a unified framework. +- **Multi-Task Requirements:** If your project needs to evolve from simple detection to [pose estimation](https://docs.ultralytics.com/tasks/pose) or segmentation using a unified framework. !!! info "Exploring Other Architectures" @@ -122,10 +122,10 @@ YOLO26 introduces several paradigm-shifting features: - **MuSGD Optimizer:** Inspired by Moonshot AI's Kimi K2, this revolutionary optimizer merges the stability of standard SGD with the accelerated momentum of Muon, bringing advanced LLM training innovations directly into computer vision. - **Enhanced CPU Speed:** By strategically removing the Distribution Focal Loss (DFL), YOLO26 achieves up to **43% faster CPU inference**, making it the undisputed champion for edge and low-power IoT device deployment. - **ProgLoss + STAL:** These advanced loss functions yield massive improvements in small-object recognition, which is critical for aerial imagery and precision robotics. -- **Task-Specific Improvements:** Featuring Semantic segmentation loss for mask generation, Residual Log-Likelihood Estimation (RLE) for Pose tracking, and specialized angle loss to resolve tricky [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) boundary issues. +- **Task-Specific Improvements:** Featuring Semantic segmentation loss for mask generation, Residual Log-Likelihood Estimation (RLE) for Pose tracking, and specialized angle loss to resolve tricky [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) boundary issues. ## Conclusion Both YOLOv5 and YOLOv7 offer robust solutions for real-time object detection. YOLOv7 remains a strong choice for raw accuracy on high-compute hardware, while YOLOv5 stands out as the ultimate developer-friendly tool, offering an exceptional balance of speed, efficiency, and a world-class ecosystem. -However, for developers looking to future-proof their pipelines and achieve the ultimate combination of speed, simplicity, and state-of-the-art accuracy, we highly recommend migrating to [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/). It encapsulates the legendary ease-of-use of the Ultralytics platform while delivering groundbreaking architectural innovations. +However, for developers looking to future-proof their pipelines and achieve the ultimate combination of speed, simplicity, and state-of-the-art accuracy, we highly recommend migrating to [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26). It encapsulates the legendary ease-of-use of the Ultralytics platform while delivering groundbreaking architectural innovations. diff --git a/docs/en/compare/yolov7-vs-yolov6.md b/docs/en/compare/yolov7-vs-yolov6.md index b18014e94ac..db1bbb6ad13 100644 --- a/docs/en/compare/yolov7-vs-yolov6.md +++ b/docs/en/compare/yolov7-vs-yolov6.md @@ -22,15 +22,15 @@ Released in mid-2022, YOLOv7 introduced several innovative strategies to optimiz - Date: 2022-07-06 - Arxiv: [2207.02696](https://arxiv.org/abs/2207.02696) - GitHub: [WongKinYiu/yolov7](https://github.com/WongKinYiu/yolov7) -- Docs: [Ultralytics YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7/) +- Docs: [Ultralytics YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7) ### Architecture Highlights YOLOv7 is characterized by its Extended Efficient Layer Aggregation Network (E-ELAN). This architecture allows the model to learn more diverse features by controlling the shortest longest gradient path. Furthermore, YOLOv7 utilizes structural re-parameterization techniques during inference to merge convolution layers, effectively reducing the parameter count and computation time without sacrificing the learned representations. -The model also features a unique auxiliary head training strategy. By using a "lead head" for final predictions and an "auxiliary head" to guide training in the middle layers, YOLOv7 achieves better convergence and richer feature extraction, particularly beneficial when tackling challenging [object detection](https://docs.ultralytics.com/tasks/detect/) tasks. +The model also features a unique auxiliary head training strategy. By using a "lead head" for final predictions and an "auxiliary head" to guide training in the middle layers, YOLOv7 achieves better convergence and richer feature extraction, particularly beneficial when tackling challenging [object detection](https://docs.ultralytics.com/tasks/detect) tasks. -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ## YOLOv6-3.0: Industrial-Grade Throughput @@ -41,7 +41,7 @@ Developed by the Meituan Vision AI Department, YOLOv6-3.0 was explicitly designe - Date: 2023-01-13 - Arxiv: [2301.05586](https://arxiv.org/abs/2301.05586) - GitHub: [meituan/YOLOv6](https://github.com/meituan/YOLOv6) -- Docs: [Ultralytics YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6/) +- Docs: [Ultralytics YOLOv6 Documentation](https://docs.ultralytics.com/models/yolov6) ### Architecture Highlights @@ -49,7 +49,7 @@ YOLOv6-3.0 adopts an EfficientRep backbone, which is highly optimized for parall Additionally, YOLOv6-3.0 utilizes an Anchor-Aided Training (AAT) strategy. This innovative approach combines the benefits of anchor-based training with anchor-free inference, allowing the model to enjoy the stability of anchors during the learning phase while maintaining the speed and simplicity of an anchor-free design during deployment. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Performance Comparison @@ -67,7 +67,7 @@ When evaluating models for production, balancing accuracy (mAP) with inference s !!! info "Hardware Considerations" - YOLOv6-3.0 is exceptionally well-suited for high-throughput GPU environments (like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/)), while YOLOv7 provides a robust balance for systems where feature retention is heavily prioritized. + YOLOv6-3.0 is exceptionally well-suited for high-throughput GPU environments (like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt)), while YOLOv7 provides a robust balance for systems where feature retention is heavily prioritized. ## The Ultralytics Advantage @@ -75,12 +75,12 @@ While the standalone repositories for YOLOv7 and YOLOv6-3.0 are powerful, levera - **Ease of Use:** Gone are the days of complex setup scripts. The Ultralytics API allows you to load, train, and deploy YOLOv7 or YOLOv6 models with minimal boilerplate code. You can easily switch between architectures by merely changing the model weights file. - **Well-Maintained Ecosystem:** Ultralytics provides a robust environment with frequent updates, ensuring native compatibility with the latest [PyTorch](https://pytorch.org/) distributions and CUDA versions. -- **Training Efficiency:** Training pipelines are deeply optimized to utilize GPU resources effectively. Furthermore, Ultralytics YOLO models generally have lower memory requirements during training compared to heavy transformer-based models (like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/)), enabling larger [batch sizes](https://www.ultralytics.com/glossary/batch-size) on consumer-grade hardware. -- **Versatility:** In addition to standard bounding box detection, the Ultralytics framework seamlessly supports advanced tasks like [pose estimation](https://docs.ultralytics.com/tasks/pose/) and [instance segmentation](https://docs.ultralytics.com/tasks/segment/) across compatible model families, a feature often lacking in isolated research repositories. +- **Training Efficiency:** Training pipelines are deeply optimized to utilize GPU resources effectively. Furthermore, Ultralytics YOLO models generally have lower memory requirements during training compared to heavy transformer-based models (like [RT-DETR](https://docs.ultralytics.com/models/rtdetr)), enabling larger [batch sizes](https://www.ultralytics.com/glossary/batch-size) on consumer-grade hardware. +- **Versatility:** In addition to standard bounding box detection, the Ultralytics framework seamlessly supports advanced tasks like [pose estimation](https://docs.ultralytics.com/tasks/pose) and [instance segmentation](https://docs.ultralytics.com/tasks/segment) across compatible model families, a feature often lacking in isolated research repositories. ### Code Example: Training and Inference -Integrating these models into your Python pipeline is straightforward. Ensure your dataset is formatted correctly (e.g., standard [COCO](https://docs.ultralytics.com/datasets/detect/coco/)) and run the following: +Integrating these models into your Python pipeline is straightforward. Ensure your dataset is formatted correctly (e.g., standard [COCO](https://docs.ultralytics.com/datasets/detect/coco)) and run the following: ```python from ultralytics import YOLO @@ -120,9 +120,9 @@ While YOLOv7 and YOLOv6-3.0 are highly capable, the rapid pace of artificial int If you are starting a new project, **YOLO26 is strongly recommended** over previous generations. It introduces several groundbreaking features: -- **End-to-End NMS-Free Design:** Building on the foundations laid by [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 natively eliminates Non-Maximum Suppression (NMS). This reduces post-processing overhead, simplifying deployment to mobile applications and ensuring highly deterministic, low-latency inference. +- **End-to-End NMS-Free Design:** Building on the foundations laid by [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 natively eliminates Non-Maximum Suppression (NMS). This reduces post-processing overhead, simplifying deployment to mobile applications and ensuring highly deterministic, low-latency inference. - **MuSGD Optimizer:** Inspired by advanced LLM training techniques (such as those used in Moonshot AI's Kimi K2), YOLO26 utilizes a hybrid optimizer combining SGD and Muon. This guarantees more stable training dynamics and drastically faster convergence. -- **Up to 43% Faster CPU Inference:** By strategically removing the Distribution Focal Loss (DFL), YOLO26 achieves massive speedups on CPUs. This makes it the undisputed champion for edge environments like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) and remote IoT sensors. +- **Up to 43% Faster CPU Inference:** By strategically removing the Distribution Focal Loss (DFL), YOLO26 achieves massive speedups on CPUs. This makes it the undisputed champion for edge environments like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) and remote IoT sensors. - **ProgLoss + STAL:** Advanced loss functions specifically engineered to improve small-object recognition, a historic weakness of single-stage detectors. -By combining these innovations with the powerful [Ultralytics Platform](https://docs.ultralytics.com/platform/), YOLO26 offers unparalleled performance, versatility, and ease of deployment for the modern machine learning engineer. +By combining these innovations with the powerful [Ultralytics Platform](https://docs.ultralytics.com/platform), YOLO26 offers unparalleled performance, versatility, and ease of deployment for the modern machine learning engineer. diff --git a/docs/en/compare/yolov7-vs-yolov8.md b/docs/en/compare/yolov7-vs-yolov8.md index 6bf9591cfe3..d981a696c1a 100644 --- a/docs/en/compare/yolov7-vs-yolov8.md +++ b/docs/en/compare/yolov7-vs-yolov8.md @@ -6,7 +6,7 @@ keywords: YOLOv7, YOLOv8, object detection, model comparison, computer vision, r # YOLOv7 vs YOLOv8: A Technical Comparison of Real-Time Detectors -The rapid evolution of computer vision has produced an array of powerful tools for developers and researchers. When deciding on the right architecture for an [object detection](https://docs.ultralytics.com/tasks/detect/) pipeline, comparing established models is essential. This technical guide provides a deep dive into the architectures, performance metrics, and ideal use cases of two highly influential models: YOLOv7 and Ultralytics YOLOv8. +The rapid evolution of computer vision has produced an array of powerful tools for developers and researchers. When deciding on the right architecture for an [object detection](https://docs.ultralytics.com/tasks/detect) pipeline, comparing established models is essential. This technical guide provides a deep dive into the architectures, performance metrics, and ideal use cases of two highly influential models: YOLOv7 and Ultralytics YOLOv8. @@ -26,7 +26,7 @@ Introduced in mid-2022, YOLOv7 focused heavily on architectural gradient path op - Date: 2022-07-06 - Arxiv: [2207.02696](https://arxiv.org/abs/2207.02696) - GitHub: [WongKinYiu/yolov7](https://github.com/WongKinYiu/yolov7) -- Docs: [Ultralytics YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7/) +- Docs: [Ultralytics YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7) **Architecture Highlights:** YOLOv7 primarily utilizes an anchor-based detection head (though it experimented with anchor-free branches) and introduces Extended Efficient Layer Aggregation Networks (E-ELAN). This design improves the learning ability of the network without destroying the original gradient path. It performs exceptionally well on server-grade [GPUs](https://en.wikipedia.org/wiki/Graphics_processing_unit), making it highly suitable for heavy-duty video analytics. @@ -34,7 +34,7 @@ YOLOv7 primarily utilizes an anchor-based detection head (though it experimented **Strengths and Weaknesses:** While YOLOv7 achieves excellent latency on dedicated hardware, its ecosystem is highly fragmented. Training requires complex command-line arguments, manual repository cloning, and strict dependency management in [PyTorch](https://pytorch.org/). Furthermore, memory requirements during training can be prohibitive on consumer hardware. -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ### Ultralytics YOLOv8: The Versatile Standard @@ -50,7 +50,7 @@ Released in early 2023, YOLOv8 completely redefined the developer experience, fo YOLOv8 introduced a natively **anchor-free** detection head, eliminating the need to manually configure anchor boxes based on the [MS COCO dataset](https://cocodataset.org/) or custom data distributions. It incorporates the C2f module to improve gradient flow and uses a decoupled head structure that separates objectness, classification, and regression tasks. This heavily accelerates convergence and boosts accuracy. **Strengths and Weaknesses:** -YOLOv8 boasts exceptional **Memory Requirements** efficiency. It requires significantly less CUDA memory during training compared to YOLOv7 and heavier transformer models, allowing developers to use larger batch sizes. Its primary strength lies in its **Versatility**, natively supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). The only minor drawback is that extremely specialized legacy pipelines built exclusively for YOLOv7 tensors might require a brief refactoring period. +YOLOv8 boasts exceptional **Memory Requirements** efficiency. It requires significantly less CUDA memory during training compared to YOLOv7 and heavier transformer models, allowing developers to use larger batch sizes. Its primary strength lies in its **Versatility**, natively supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). The only minor drawback is that extremely specialized legacy pipelines built exclusively for YOLOv7 tensors might require a brief refactoring period. [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -79,7 +79,7 @@ _Note: YOLOv8x achieves the highest mAP in this grouping, while YOLOv8n dominate When it comes to **Ease of Use**, Ultralytics YOLOv8 operates in a league of its own. Older architectures like YOLOv7 require cloning specific repositories and running verbose command-line scripts to configure datasets and paths. -Conversely, YOLOv8's `ultralytics` package offers a highly streamlined developer experience. **Training Efficiency** is maximized through automatic data downloading, ready-to-use pretrained weights, and seamless [exporting capabilities](https://docs.ultralytics.com/modes/export/) to formats like [ONNX](https://onnx.ai/) and [TensorRT](https://developer.nvidia.com/tensorrt). +Conversely, YOLOv8's `ultralytics` package offers a highly streamlined developer experience. **Training Efficiency** is maximized through automatic data downloading, ready-to-use pretrained weights, and seamless [exporting capabilities](https://docs.ultralytics.com/modes/export) to formats like [ONNX](https://onnx.ai/) and [TensorRT](https://developer.nvidia.com/tensorrt). Here is how easily you can load, train, and run inference using the Ultralytics Python API: @@ -101,7 +101,7 @@ predictions[0].show() !!! note "Experiment Tracking" - YOLOv8 integrates natively with popular MLops tools like [Weights & Biases](https://wandb.ai/site) and [ClearML](https://docs.ultralytics.com/integrations/clearml/), allowing you to monitor your [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/) and training metrics in real-time. + YOLOv8 integrates natively with popular MLops tools like [Weights & Biases](https://wandb.ai/site) and [ClearML](https://docs.ultralytics.com/integrations/clearml), allowing you to monitor your [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning) and training metrics in real-time. ## Ideal Use Cases @@ -115,7 +115,7 @@ Choosing between these architectures often comes down to the specific constraint ### When to Choose YOLOv8 - **Cross-Platform Production:** Ideal for teams that need to deploy seamlessly across cloud GPUs, mobile devices, and browsers. -- **Multi-Task Requirements:** If your project needs to move beyond bounding boxes and leverage rich [instance segmentation masks](https://docs.ultralytics.com/tasks/segment/) or [pose keypoints](https://docs.ultralytics.com/tasks/pose/). +- **Multi-Task Requirements:** If your project needs to move beyond bounding boxes and leverage rich [instance segmentation masks](https://docs.ultralytics.com/tasks/segment) or [pose keypoints](https://docs.ultralytics.com/tasks/pose). - **Resource-Constrained Edge:** YOLOv8 Nano (`yolov8n`) provides incredible accuracy-to-speed ratios for robotics, drones, and IoT sensors. ## Looking Forward: The Generational Leap to YOLO26 @@ -125,9 +125,9 @@ While YOLOv8 remains a highly robust choice, the field of computer vision moves Released in January 2026, [YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) pushes the boundaries of what is possible on edge devices: - **End-to-End NMS-Free Design:** YOLO26 is natively end-to-end, completely eliminating Non-Maximum Suppression (NMS) post-processing. This ensures significantly faster, simpler deployment pipelines without the latency bottlenecks of traditional dense prediction models. -- **DFL Removal:** By removing Distribution Focal Loss, YOLO26 achieves much simpler [model deployment options](https://docs.ultralytics.com/guides/model-deployment-options/) and superior edge compatibility. +- **DFL Removal:** By removing Distribution Focal Loss, YOLO26 achieves much simpler [model deployment options](https://docs.ultralytics.com/guides/model-deployment-options) and superior edge compatibility. - **Up to 43% Faster CPU Inference:** Heavily optimized for constrained environments like Raspberry Pi and embedded systems, beating all prior generations in CPU throughput. - **MuSGD Optimizer:** Inspired by Large Language Model (LLM) training paradigms, YOLO26 incorporates a hybrid of SGD and Muon. This delivers unprecedented training stability and lightning-fast convergence. - **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, which is highly critical for aerial imagery, automated agriculture, and robotics. -Whether you are scaling up to massive video analytics clusters with YOLOv8 or pushing inference to tiny edge devices with the cutting-edge YOLO26, the [Ultralytics Platform](https://docs.ultralytics.com/platform/) provides the tools to manage your entire AI lifecycle seamlessly. +Whether you are scaling up to massive video analytics clusters with YOLOv8 or pushing inference to tiny edge devices with the cutting-edge YOLO26, the [Ultralytics Platform](https://docs.ultralytics.com/platform) provides the tools to manage your entire AI lifecycle seamlessly. diff --git a/docs/en/compare/yolov7-vs-yolov9.md b/docs/en/compare/yolov7-vs-yolov9.md index 629db436ef5..f2fb3eb3c42 100644 --- a/docs/en/compare/yolov7-vs-yolov9.md +++ b/docs/en/compare/yolov7-vs-yolov9.md @@ -6,7 +6,7 @@ keywords: YOLOv7, YOLOv9, object detection, model comparison, YOLO architecture, # YOLOv7 vs YOLOv9: A Technical Deep Dive into Modern Object Detection -The landscape of real-time [object detection](https://www.ultralytics.com/glossary/object-detection) has evolved rapidly, with each new iteration pushing the boundaries of what is possible on edge devices and cloud servers alike. When evaluating architectures for computer vision projects, developers frequently compare established benchmarks with newer innovations. This comprehensive guide compares two pivotal milestones in the YOLO family: [YOLOv7](https://docs.ultralytics.com/models/yolov7/) and [YOLOv9](https://docs.ultralytics.com/models/yolov9/). +The landscape of real-time [object detection](https://www.ultralytics.com/glossary/object-detection) has evolved rapidly, with each new iteration pushing the boundaries of what is possible on edge devices and cloud servers alike. When evaluating architectures for computer vision projects, developers frequently compare established benchmarks with newer innovations. This comprehensive guide compares two pivotal milestones in the YOLO family: [YOLOv7](https://docs.ultralytics.com/models/yolov7) and [YOLOv9](https://docs.ultralytics.com/models/yolov9). We will analyze their architectural breakthroughs, performance metrics, and ideal deployment scenarios to help you choose the right model for your application. We will also explore how the [Ultralytics Platform](https://platform.ultralytics.com/explore) unifies these models, making them easier to train, validate, and deploy. @@ -31,7 +31,7 @@ Released in mid-2022, YOLOv7 established itself as a highly reliable and heavily **Architectural Innovations:** YOLOv7 features the Extended Efficient Layer Aggregation Network (E-ELAN), which allows the model to learn more diverse features by expanding, shuffling, and merging cardinality. This design results in excellent GPU utilization and [inference latency](https://www.ultralytics.com/glossary/inference-latency). However, it can require significant memory during complex training runs compared to modern iterations. -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ### YOLOv9: Solving the Information Bottleneck @@ -45,11 +45,11 @@ Introduced in early 2024 by the same research team, YOLOv9 tackles the "informat **Architectural Innovations:** YOLOv9 introduces Programmable Gradient Information (PGI) and the Generalized Efficient Layer Aggregation Network (GELAN). PGI ensures that reliable gradients are preserved and fed back to update weights accurately. GELAN maximizes parameter efficiency, enabling YOLOv9 to achieve high accuracy with significantly fewer [FLOPs](https://www.ultralytics.com/glossary/flops) than its predecessors. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ## Performance Analysis -When choosing between architectures, AI engineers must balance accuracy, [inference speed](https://docs.ultralytics.com/guides/yolo-performance-metrics/), and computational cost. The table below highlights the performance differences across these models on the standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). +When choosing between architectures, AI engineers must balance accuracy, [inference speed](https://docs.ultralytics.com/guides/yolo-performance-metrics), and computational cost. The table below highlights the performance differences across these models on the standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -65,7 +65,7 @@ When choosing between architectures, AI engineers must balance accuracy, [infere ### Key Takeaways - **Parameter Efficiency:** YOLOv9m matches the accuracy of YOLOv7l (51.4% mAP) while utilizing nearly **45% fewer parameters** (20.0M vs 36.9M). This drastic reduction makes YOLOv9m much easier to deploy on memory-constrained [edge AI](https://www.ultralytics.com/glossary/edge-ai) devices. -- **Micro-Deployments:** The introduction of the YOLOv9t (tiny) variant provides incredible speeds (2.3ms on T4 [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/)) for environments where real-time constraints are absolute. +- **Micro-Deployments:** The introduction of the YOLOv9t (tiny) variant provides incredible speeds (2.3ms on T4 [TensorRT](https://docs.ultralytics.com/integrations/tensorrt)) for environments where real-time constraints are absolute. - **Maximum Accuracy:** For applications where precision is paramount, YOLOv9e pushes detection accuracy to 55.6% mAP, significantly outperforming YOLOv7x. !!! tip "Future-Proofing Your Computer Vision Projects" @@ -74,13 +74,13 @@ When choosing between architectures, AI engineers must balance accuracy, [infere ## The Ultralytics Advantage -Choosing a model architecture is only the first step. The software ecosystem surrounding the model determines how quickly you can move from prototype to production. Integrating these models through the [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) provides substantial benefits for developers and researchers. +Choosing a model architecture is only the first step. The software ecosystem surrounding the model determines how quickly you can move from prototype to production. Integrating these models through the [Ultralytics Python API](https://docs.ultralytics.com/usage/python) provides substantial benefits for developers and researchers. ### Ease of Use and Training Efficiency -Historically, training YOLOv7 required complex data preparation and heavily customized scripts. The Ultralytics framework abstracts away these deep learning complexities. Developers can easily switch between architectures, experiment with [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/), and utilize intelligent [data augmentation](https://docs.ultralytics.com/reference/data/augment/) pipelines with minimal code. +Historically, training YOLOv7 required complex data preparation and heavily customized scripts. The Ultralytics framework abstracts away these deep learning complexities. Developers can easily switch between architectures, experiment with [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning), and utilize intelligent [data augmentation](https://docs.ultralytics.com/reference/data/augment) pipelines with minimal code. -Furthermore, Ultralytics optimizes [memory usage](https://www.ultralytics.com/glossary/batch-size) during training and inference. Unlike heavy [transformer models](https://www.ultralytics.com/glossary/transformer) (such as [RT-DETR](https://docs.ultralytics.com/models/rtdetr/)), Ultralytics YOLO architectures train significantly faster and require much less CUDA memory, making them ideal for consumer-grade GPUs. +Furthermore, Ultralytics optimizes [memory usage](https://www.ultralytics.com/glossary/batch-size) during training and inference. Unlike heavy [transformer models](https://www.ultralytics.com/glossary/transformer) (such as [RT-DETR](https://docs.ultralytics.com/models/rtdetr)), Ultralytics YOLO architectures train significantly faster and require much less CUDA memory, making them ideal for consumer-grade GPUs. ### Code Example: Streamlined Training @@ -110,7 +110,7 @@ model.export(format="onnx") ### Unmatched Versatility Across Tasks -A well-maintained ecosystem means access to diverse computer vision tasks. While YOLOv7 was primarily built for object detection (with later experimental forks for other tasks), modern Ultralytics models are natively built for versatility. Out of the box, you can perform [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection seamlessly. +A well-maintained ecosystem means access to diverse computer vision tasks. While YOLOv7 was primarily built for object detection (with later experimental forks for other tasks), modern Ultralytics models are natively built for versatility. Out of the box, you can perform [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection seamlessly. ## Ideal Use Cases and Applications diff --git a/docs/en/compare/yolov7-vs-yolox.md b/docs/en/compare/yolov7-vs-yolox.md index 425f42f1152..63d200ea846 100644 --- a/docs/en/compare/yolov7-vs-yolox.md +++ b/docs/en/compare/yolov7-vs-yolox.md @@ -26,9 +26,9 @@ Developed by the researchers who maintained the CSPNet and Scaled-YOLOv4 archite - **Date:** 2022-07-06 - **Arxiv:** [https://arxiv.org/abs/2207.02696](https://arxiv.org/abs/2207.02696) - **GitHub:** [https://github.com/WongKinYiu/yolov7](https://github.com/WongKinYiu/yolov7) -- **Docs:** [Ultralytics YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7/) +- **Docs:** [Ultralytics YOLOv7 Documentation](https://docs.ultralytics.com/models/yolov7) -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ### YOLOX Details @@ -86,8 +86,8 @@ A crucial factor for researchers and developers is the ease of implementation. H Today, the most effective way to utilize these architectures is through the well-maintained Ultralytics ecosystem. Ultralytics provides a unified, highly intuitive Python API that drastically simplifies training, validation, and deployment. - **Ease of Use:** With just a few lines of code, you can initiate a training loop, mitigating the steep learning curve associated with raw PyTorch implementations. -- **Training Efficiency:** Ultralytics YOLO models inherently utilize lower memory during training compared to heavy transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). This allows developers to maximize batch sizes on consumer hardware. -- **Versatility:** Beyond simple bounding boxes, the ecosystem effortlessly extends to tasks like [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/) and [Pose Estimation](https://docs.ultralytics.com/tasks/pose/). +- **Training Efficiency:** Ultralytics YOLO models inherently utilize lower memory during training compared to heavy transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). This allows developers to maximize batch sizes on consumer hardware. +- **Versatility:** Beyond simple bounding boxes, the ecosystem effortlessly extends to tasks like [Instance Segmentation](https://docs.ultralytics.com/tasks/segment) and [Pose Estimation](https://docs.ultralytics.com/tasks/pose). Here is a 100% runnable example demonstrating how to train a model utilizing the Ultralytics API: @@ -110,13 +110,13 @@ results = model.train( model.export(format="onnx") ``` -By standardizing the [export pipeline](https://docs.ultralytics.com/modes/export/), developers can effortlessly transition their weights to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) or [ONNX](https://docs.ultralytics.com/integrations/onnx/), ensuring high-speed inference on target hardware. +By standardizing the [export pipeline](https://docs.ultralytics.com/modes/export), developers can effortlessly transition their weights to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) or [ONNX](https://docs.ultralytics.com/integrations/onnx), ensuring high-speed inference on target hardware. ## Ideal Use Cases and Real-World Applications Choosing between YOLOX and YOLOv7 largely depends on deployment targets: -- **YOLOX for Edge AI:** The YOLOX-Nano and YOLOX-Tiny variants are highly suitable for deployment on low-power devices. If you are building a smart security camera on a [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/), the simple anchor-free convolutions of YOLOX translate easily to edge accelerators. +- **YOLOX for Edge AI:** The YOLOX-Nano and YOLOX-Tiny variants are highly suitable for deployment on low-power devices. If you are building a smart security camera on a [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi), the simple anchor-free convolutions of YOLOX translate easily to edge accelerators. - **YOLOv7 for High-Fidelity Analytics:** If you are processing high-resolution satellite imagery or executing complex [manufacturing quality control](https://www.ultralytics.com/blog/improving-manufacturing-with-computer-vision), the high mAP of YOLOv7x, powered by high-end NVIDIA GPUs, ensures that even the smallest anomalies are detected. ## The Future: Upgrading to Ultralytics YOLO26 @@ -125,13 +125,13 @@ While YOLOv7 and YOLOX were groundbreaking at their inception, the computer visi Here is why upgrading is highly recommended: -- **End-to-End NMS-Free Design:** YOLO26 natively eliminates Non-Maximum Suppression (NMS) during post-processing. Pioneered initially in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), this ensures consistently low latency, simplifying deployment on devices lacking NMS hardware support. +- **End-to-End NMS-Free Design:** YOLO26 natively eliminates Non-Maximum Suppression (NMS) during post-processing. Pioneered initially in [YOLOv10](https://docs.ultralytics.com/models/yolov10), this ensures consistently low latency, simplifying deployment on devices lacking NMS hardware support. - **DFL Removal:** By removing Distribution Focal Loss, YOLO26 achieves vastly better compatibility with low-power edge devices and straightforward ONNX exports. - **MuSGD Optimizer:** Inspired by LLM training innovations, YOLO26 leverages a hybrid MuSGD optimizer, ensuring faster convergence and incredibly stable training dynamics. - **Up to 43% Faster CPU Inference:** Optimized heavily for real-world hardware, YOLO26 thrives on standard CPUs without requiring expensive GPU infrastructure. -- **ProgLoss + STAL:** These advanced loss functions drastically improve small-object recognition, a critical feature for [aerial drone inspections](https://docs.ultralytics.com/datasets/detect/visdrone/) and sophisticated IoT networks. +- **ProgLoss + STAL:** These advanced loss functions drastically improve small-object recognition, a critical feature for [aerial drone inspections](https://docs.ultralytics.com/datasets/detect/visdrone) and sophisticated IoT networks. -For developers seeking the best performance balance across [object detection](https://docs.ultralytics.com/tasks/detect/), segmentation, and beyond, deploying models via the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolo26) provides an unparalleled, zero-friction experience. +For developers seeking the best performance balance across [object detection](https://docs.ultralytics.com/tasks/detect), segmentation, and beyond, deploying models via the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolo26) provides an unparalleled, zero-friction experience. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } diff --git a/docs/en/compare/yolov8-vs-damo-yolo.md b/docs/en/compare/yolov8-vs-damo-yolo.md index 7c2fd3c9374..d232bb4b697 100644 --- a/docs/en/compare/yolov8-vs-damo-yolo.md +++ b/docs/en/compare/yolov8-vs-damo-yolo.md @@ -23,7 +23,7 @@ Both models were introduced around the same time but stem from different design - Organization: [Ultralytics](https://www.ultralytics.com/) - Date: 2023-01-10 - GitHub: [Ultralytics GitHub Repository](https://github.com/ultralytics/ultralytics) -- Docs: [YOLOv8 Official Documentation](https://docs.ultralytics.com/models/yolov8/) +- Docs: [YOLOv8 Official Documentation](https://docs.ultralytics.com/models/yolov8) [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -43,7 +43,7 @@ Both models were introduced around the same time but stem from different design [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) introduced significant improvements over its predecessors, cementing its status as a highly reliable state-of-the-art model. It features an anchor-free detection head, which reduces the number of box predictions and speeds up inference. The architecture utilizes a decoupled head, separating objectness, classification, and regression tasks, leading to more accurate bounding box predictions. -Furthermore, YOLOv8 implements [Distribution Focal Loss (DFL)](https://docs.ultralytics.com/reference/utils/loss/) alongside CIoU loss, enhancing the model's ability to precisely localize object boundaries, especially for smaller or occluded targets. Its streamlined backbone is highly optimized for both GPU and CPU execution. +Furthermore, YOLOv8 implements [Distribution Focal Loss (DFL)](https://docs.ultralytics.com/reference/utils/loss) alongside CIoU loss, enhancing the model's ability to precisely localize object boundaries, especially for smaller or occluded targets. Its streamlined backbone is highly optimized for both GPU and CPU execution. ### DAMO-YOLO: Driven by Architecture Search @@ -99,7 +99,7 @@ path = model.export(format="onnx") ### Versatility Across Vision Tasks -DAMO-YOLO is strictly built for bounding-box object detection. In contrast, the YOLOv8 architecture natively supports multiple tasks. By simply swapping the model weights, developers can perform [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), and [Pose Estimation](https://docs.ultralytics.com/tasks/pose/) without changing their underlying deployment codebase. This versatility makes Ultralytics models much more practical for complex applications. +DAMO-YOLO is strictly built for bounding-box object detection. In contrast, the YOLOv8 architecture natively supports multiple tasks. By simply swapping the model weights, developers can perform [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Image Classification](https://docs.ultralytics.com/tasks/classify), and [Pose Estimation](https://docs.ultralytics.com/tasks/pose) without changing their underlying deployment codebase. This versatility makes Ultralytics models much more practical for complex applications. ## Real-World Use Cases @@ -107,10 +107,10 @@ DAMO-YOLO is strictly built for bounding-box object detection. In contrast, the YOLOv8's combination of speed, accuracy, and ease of deployment makes it ideal for: -- **Smart Retail Analytics:** Performing [object tracking](https://docs.ultralytics.com/modes/track/) to monitor customer behavior or automate inventory checks. +- **Smart Retail Analytics:** Performing [object tracking](https://docs.ultralytics.com/modes/track) to monitor customer behavior or automate inventory checks. - **Agricultural Robotics:** Leveraging its strong performance on varied hardware to identify crops or pests in real-time. - **Healthcare Diagnostics:** Using instance segmentation to map anomalies in medical imagery quickly and accurately. -- **Edge Deployments:** The seamless integration with export formats like [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) and [CoreML](https://docs.ultralytics.com/integrations/coreml/) allows YOLOv8 to shine on constrained devices. +- **Edge Deployments:** The seamless integration with export formats like [OpenVINO](https://docs.ultralytics.com/integrations/openvino) and [CoreML](https://docs.ultralytics.com/integrations/coreml) allows YOLOv8 to shine on constrained devices. ### When to use DAMO-YOLO @@ -127,7 +127,7 @@ Choosing between YOLOv8 and DAMO-YOLO depends on your specific project requireme YOLOv8 is a strong choice for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. @@ -141,11 +141,11 @@ DAMO-YOLO is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Looking Forward: Newer Ultralytics Models diff --git a/docs/en/compare/yolov8-vs-efficientdet.md b/docs/en/compare/yolov8-vs-efficientdet.md index d9a6ab63a6e..53e8d35f9e6 100644 --- a/docs/en/compare/yolov8-vs-efficientdet.md +++ b/docs/en/compare/yolov8-vs-efficientdet.md @@ -21,7 +21,7 @@ Both models approach the challenge of identifying and localizing objects in an i ### Ultralytics YOLOv8 -Released by Ultralytics in January 2023, YOLOv8 represented a major leap forward in the YOLO family line. Authored by Glenn Jocher, Ayush Chaurasia, and Jing Qiu, it was designed from the ground up to support multiple vision tasks seamlessly, including [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and image classification. +Released by Ultralytics in January 2023, YOLOv8 represented a major leap forward in the YOLO family line. Authored by Glenn Jocher, Ayush Chaurasia, and Jing Qiu, it was designed from the ground up to support multiple vision tasks seamlessly, including [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and image classification. The architecture introduces an anchor-free detection head, which heavily reduces the number of box predictions and speeds up Non-Maximum Suppression (NMS). Its backbone utilizes a novel **C2f module** (Cross-Stage Partial bottleneck with two convolutions) to improve gradient flow during training while maintaining a lightweight footprint. This makes YOLOv8 exceptionally efficient when compiled to formats like [NVIDIA TensorRT](https://developer.nvidia.com/tensorrt) or [ONNX](https://onnx.ai/). @@ -29,7 +29,7 @@ The architecture introduces an anchor-free detection head, which heavily reduces ### EfficientDet -Authored by Mingxing Tan, Ruoming Pang, and Quoc V. Le at Google and released in late 2019, EfficientDet focuses on scalable efficiency. Described in their [official Arxiv paper](https://arxiv.org/abs/1911.09070), the model heavily leverages the [AutoML ecosystem](https://cloud.google.com/automl). +Authored by Mingxing Tan, Ruoming Pang, and Quoc V. Le at Google and released in late 2019, EfficientDet focuses on scalable efficiency. Described in their [official Arxiv paper](https://arxiv.org/abs/1911.09070), the model heavily leverages the [AutoML ecosystem](https://cloud.google.com/products/gemini-enterprise-agent-platform). The defining characteristic of EfficientDet is its **Bi-directional Feature Pyramid Network (BiFPN)**, which enables easy and fast multi-scale feature fusion. Combined with an EfficientNet backbone, the architecture uses a compound scaling method that uniformly scales the resolution, depth, and width for all backbone, feature network, and box/class prediction networks at the same time. While this results in excellent parameter efficiency, the complex network topology often struggles to achieve optimal real-time speeds on standard GPUs. @@ -37,7 +37,7 @@ The defining characteristic of EfficientDet is its **Bi-directional Feature Pyra ## Performance and Metrics Comparison -When comparing object detectors, [mean Average Precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics/) and inference latency are the primary benchmarks. The table below illustrates how the YOLOv8 variants and the EfficientDet (d0-d7) family compare across standard metrics on datasets like [COCO](https://cocodataset.org/). +When comparing object detectors, [mean Average Precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics) and inference latency are the primary benchmarks. The table below illustrates how the YOLOv8 variants and the EfficientDet (d0-d7) family compare across standard metrics on datasets like [COCO](https://cocodataset.org/). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | --------------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -68,7 +68,7 @@ EfficientDet relies heavily on [TensorFlow](https://www.tensorflow.org/) and spe In contrast, **Ultralytics YOLOv8** is built natively on [PyTorch](https://pytorch.org/), offering unmatched ease of use. Developers can initiate complex training loops with a single line of Python code or CLI command. Furthermore, the model memory requirements during training are heavily optimized; YOLOv8 allows developers with modest consumer GPUs to train robust models without encountering out-of-memory (OOM) errors that frequently plague transformer-heavy architectures. -The seamless integration with the [Ultralytics Platform](https://platform.ultralytics.com) takes this a step further, providing a no-code interface for dataset annotation, model training, and one-click cloud deployment. Features like automatic [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/) ensure that you always get the best possible accuracy for your custom datasets. +The seamless integration with the [Ultralytics Platform](https://platform.ultralytics.com) takes this a step further, providing a no-code interface for dataset annotation, model training, and one-click cloud deployment. Features like automatic [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning) ensure that you always get the best possible accuracy for your custom datasets. ### Python Code Example: YOLOv8 Inference @@ -101,14 +101,14 @@ Furthermore, YOLO26 incorporates several groundbreaking training innovations: - **MuSGD Optimizer:** Inspired by advanced LLM training techniques, this hybrid of SGD and Muon ensures highly stable training and vastly accelerated convergence rates. - **Up to 43% Faster CPU Inference:** Thanks to the NMS removal and a heavily optimized backbone, YOLO26 achieves unprecedented speeds on CPU-only edge devices without relying on dedicated NPUs. - **ProgLoss + STAL:** These advanced loss functions deliver a notable leap in small-object recognition accuracy, making YOLO26 indispensable for aerial imagery and precision IoT sensors. -- **DFL Removal:** The Distribution Focal Loss has been completely removed to drastically simplify the export process to formats like [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) and CoreML. +- **DFL Removal:** The Distribution Focal Loss has been completely removed to drastically simplify the export process to formats like [OpenVINO](https://docs.ultralytics.com/integrations/openvino) and CoreML. ## Use Cases and Recommendations Selecting between these architectures ultimately depends on your deployment constraints and legacy requirements. -- **Choose Ultralytics YOLOv8 if:** You are building modern, versatile computer vision applications that demand high accuracy, real-time GPU inference, and a frictionless developer experience. Its strong performance across [classification, segmentation, and detection tasks](https://docs.ultralytics.com/tasks/) makes it a powerful multi-tool for retail analytics, robotics, and security systems. +- **Choose Ultralytics YOLOv8 if:** You are building modern, versatile computer vision applications that demand high accuracy, real-time GPU inference, and a frictionless developer experience. Its strong performance across [classification, segmentation, and detection tasks](https://docs.ultralytics.com/tasks) makes it a powerful multi-tool for retail analytics, robotics, and security systems. - **Choose EfficientDet if:** You are locked into legacy TensorFlow workflows and your primary concern is minimizing parameter counts and theoretical FLOPs, perhaps for research purposes rather than strict real-time industrial deployment. - **Choose Ultralytics YOLO26 if:** You are starting a new project and require the absolute best. Its native end-to-end NMS-free architecture makes it the ultimate choice for both ultra-fast edge deployments and heavy cloud processing. -If you are exploring other highly capable frameworks within the Ultralytics ecosystem, you may also consider [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) for balanced legacy performance or [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) for a transformer-based approach to real-time detection. +If you are exploring other highly capable frameworks within the Ultralytics ecosystem, you may also consider [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) for balanced legacy performance or [RT-DETR](https://docs.ultralytics.com/models/rtdetr) for a transformer-based approach to real-time detection. diff --git a/docs/en/compare/yolov8-vs-pp-yoloe.md b/docs/en/compare/yolov8-vs-pp-yoloe.md index 3817620dd86..2481c1af73d 100644 --- a/docs/en/compare/yolov8-vs-pp-yoloe.md +++ b/docs/en/compare/yolov8-vs-pp-yoloe.md @@ -21,7 +21,7 @@ Introduced by Ultralytics, YOLOv8 quickly established itself as a cornerstone fo - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2023-01-10 - **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8/) +- **Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8) [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -29,7 +29,7 @@ Introduced by Ultralytics, YOLOv8 quickly established itself as a cornerstone fo YOLOv8 features a highly optimized anchor-free design and incorporates a decoupled head to independently process objectness, classification, and regression tasks. This structural refinement leads to better feature representation and faster convergence during training. -Unlike many specialized models, YOLOv8 offers unmatched versatility. Beyond bounding box detection, the same unified architecture and API natively support [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +Unlike many specialized models, YOLOv8 offers unmatched versatility. Beyond bounding box detection, the same unified architecture and API natively support [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb). !!! tip "Streamlined Development" @@ -82,7 +82,7 @@ Furthermore, YOLO models heavily outperform large transformer-based architecture The true defining factor between these architectures lies in the user experience. -The **[Ultralytics Platform](https://platform.ultralytics.com)** offers a well-maintained ecosystem that abstracts away the friction of machine learning operations. It provides an incredibly simple API, extensive documentation, and native tools for data logging, hyperparameter tuning, and cross-platform export. Whether you need to deploy via [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), or [CoreML](https://developer.apple.com/documentation/coreml), Ultralytics handles it seamlessly. +The **[Ultralytics Platform](https://platform.ultralytics.com)** offers a well-maintained ecosystem that abstracts away the friction of machine learning operations. It provides an incredibly simple API, extensive documentation, and native tools for data logging, hyperparameter tuning, and cross-platform export. Whether you need to deploy via [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), or [CoreML](https://developer.apple.com/documentation/coreml), Ultralytics handles it seamlessly. Conversely, PP-YOLOE+ often requires deep knowledge of the PaddlePaddle framework. Converting these models to run efficiently on standard [NVIDIA GPUs](https://www.nvidia.com/en-us/data-center/tesla-t4/) or edge devices outside of the Baidu hardware ecosystem can be a complex, multi-step process lacking the streamlined automation found in Ultralytics tools. @@ -111,7 +111,7 @@ Choosing between YOLOv8 and PP-YOLOE+ depends on your specific project requireme YOLOv8 is a strong choice for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. @@ -125,11 +125,11 @@ PP-YOLOE+ is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Looking Forward: The YOLO26 Advantage @@ -138,7 +138,7 @@ For those looking to build future-proof applications, the recently released **[U - **End-to-End NMS-Free Design:** YOLO26 natively eliminates the need for Non-Maximum Suppression post-processing, dramatically reducing latency variability and simplifying deployment logic. - **MuSGD Optimizer:** Integrating LLM training innovations into vision AI, this hybrid of SGD and [Muon](https://github.com/KellerJordan/Muon) ensures incredibly stable training dynamics and faster convergence. - **Up to 43% Faster CPU Inference:** By removing Distribution Focal Loss (DFL), YOLO26 provides unmatched speed on edge devices and standard CPUs, making it ideal for IoT and mobile applications. -- **ProgLoss + STAL:** These advanced loss functions deliver notable improvements in small-object recognition, a critical requirement for [drone analytics](https://docs.ultralytics.com/datasets/detect/visdrone/) and aerial imagery. +- **ProgLoss + STAL:** These advanced loss functions deliver notable improvements in small-object recognition, a critical requirement for [drone analytics](https://docs.ultralytics.com/datasets/detect/visdrone) and aerial imagery. !!! note "Upgrade Recommendation" diff --git a/docs/en/compare/yolov8-vs-rtdetr.md b/docs/en/compare/yolov8-vs-rtdetr.md index 5d000df9cac..9e5181021c5 100644 --- a/docs/en/compare/yolov8-vs-rtdetr.md +++ b/docs/en/compare/yolov8-vs-rtdetr.md @@ -23,17 +23,17 @@ Ultralytics YOLOv8 represents a major milestone in the YOLO (You Only Look Once) - Organization: [Ultralytics](https://www.ultralytics.com/) - Date: January 10, 2023 - GitHub: [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- Docs: [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8/) +- Docs: [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8) ### Architecture and Strengths -YOLOv8 introduces a streamlined architecture that optimizes both feature extraction and bounding box regression. It is an anchor-free detector, which simplifies the prediction head and reduces the number of hyperparameter tweaks required during training. This architecture ensures a fantastic [performance balance](https://docs.ultralytics.com/guides/yolo-performance-metrics/) between inference speed and mean average precision (mAP), making it highly suitable for real-world deployment on both edge devices and cloud servers. +YOLOv8 introduces a streamlined architecture that optimizes both feature extraction and bounding box regression. It is an anchor-free detector, which simplifies the prediction head and reduces the number of hyperparameter tweaks required during training. This architecture ensures a fantastic [performance balance](https://docs.ultralytics.com/guides/yolo-performance-metrics) between inference speed and mean average precision (mAP), making it highly suitable for real-world deployment on both edge devices and cloud servers. -Furthermore, YOLOv8 requires significantly lower [memory requirements](https://docs.ultralytics.com/guides/model-training-tips/) during training compared to transformer-based architectures. This allows developers to train models on standard consumer GPUs without encountering out-of-memory errors. +Furthermore, YOLOv8 requires significantly lower [memory requirements](https://docs.ultralytics.com/guides/model-training-tips) during training compared to transformer-based architectures. This allows developers to train models on standard consumer GPUs without encountering out-of-memory errors. ### Versatility -One of the defining strengths of YOLOv8 is its native versatility. While many models focus solely on bounding boxes, YOLOv8 provides out-of-the-box support for [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. +One of the defining strengths of YOLOv8 is its native versatility. While many models focus solely on bounding boxes, YOLOv8 provides out-of-the-box support for [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -89,7 +89,7 @@ Choosing between YOLOv8 and RT-DETR depends on your specific project requirement YOLOv8 is a strong choice for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. @@ -103,19 +103,19 @@ RT-DETR is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage -Choosing a model goes beyond raw metrics; the surrounding software ecosystem is crucial for developer productivity. The [Ultralytics ecosystem](https://docs.ultralytics.com/platform/) is renowned for its ease of use, providing a unified Python API that simplifies the entire machine learning lifecycle. +Choosing a model goes beyond raw metrics; the surrounding software ecosystem is crucial for developer productivity. The [Ultralytics ecosystem](https://docs.ultralytics.com/platform) is renowned for its ease of use, providing a unified Python API that simplifies the entire machine learning lifecycle. From dataset management to distributed training, Ultralytics abstracts away complex boilerplate code. Developers benefit from readily available pre-trained weights and seamless integration with platforms like [Hugging Face](https://huggingface.co/) and monitoring tools. This well-maintained ecosystem guarantees active development, frequent updates, and robust community support. -Furthermore, training efficiency is a hallmark of Ultralytics YOLO models. They are highly optimized for fast convergence and lower memory footprints during the [training process](https://docs.ultralytics.com/modes/train/), which significantly accelerates experimentation cycles compared to transformer-based detectors like RTDETRv2. +Furthermore, training efficiency is a hallmark of Ultralytics YOLO models. They are highly optimized for fast convergence and lower memory footprints during the [training process](https://docs.ultralytics.com/modes/train), which significantly accelerates experimentation cycles compared to transformer-based detectors like RTDETRv2. ## Looking Ahead: The Power of YOLO26 @@ -155,7 +155,7 @@ export_path = model.export(format="onnx") !!! tip "Deployment Ready" - Ultralytics supports one-click exports to numerous formats, including ONNX, TensorRT, and CoreML, simplifying [model deployment options](https://docs.ultralytics.com/guides/model-deployment-options/) across varying hardware architectures. + Ultralytics supports one-click exports to numerous formats, including ONNX, TensorRT, and CoreML, simplifying [model deployment options](https://docs.ultralytics.com/guides/model-deployment-options) across varying hardware architectures. ## Conclusion diff --git a/docs/en/compare/yolov8-vs-yolo11.md b/docs/en/compare/yolov8-vs-yolo11.md index 2960f166b64..30ee42c5e25 100644 --- a/docs/en/compare/yolov8-vs-yolo11.md +++ b/docs/en/compare/yolov8-vs-yolo11.md @@ -43,7 +43,7 @@ Building upon the success of its predecessors, YOLO11 refined the core architect !!! tip "Other Architectures" - If you are exploring alternative approaches, Ultralytics also supports transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) and zero-shot open-vocabulary detectors like [YOLO-World](https://docs.ultralytics.com/models/yolo-world/). However, for optimal latency and memory efficiency, standard YOLO architectures typically remain the preferred choice. + If you are exploring alternative approaches, Ultralytics also supports transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) and zero-shot open-vocabulary detectors like [YOLO-World](https://docs.ultralytics.com/models/yolo-world). However, for optimal latency and memory efficiency, standard YOLO architectures typically remain the preferred choice. ## Architectural and Methodological Differences @@ -100,10 +100,10 @@ The seamless integration extends to the [Ultralytics Platform](https://platform. A major hallmark of the Ultralytics framework is its inherent **versatility**. Both YOLOv8 and YOLO11 support a wide range of computer vision tasks beyond standard object detection: -- **[Instance Segmentation](https://docs.ultralytics.com/tasks/segment/):** Highly accurate pixel-level masks useful for medical imaging and autonomous driving. -- **[Pose Estimation](https://docs.ultralytics.com/tasks/pose/):** Keypoint detection tailored for sports analytics and human-computer interaction. -- **[Image Classification](https://docs.ultralytics.com/tasks/classify/):** Lightweight categorization utilizing backbones trained on [ImageNet](https://www.image-net.org/). -- **[Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/):** Critical for identifying rotated objects in satellite imagery. +- **[Instance Segmentation](https://docs.ultralytics.com/tasks/segment):** Highly accurate pixel-level masks useful for medical imaging and autonomous driving. +- **[Pose Estimation](https://docs.ultralytics.com/tasks/pose):** Keypoint detection tailored for sports analytics and human-computer interaction. +- **[Image Classification](https://docs.ultralytics.com/tasks/classify):** Lightweight categorization utilizing backbones trained on [ImageNet](https://www.image-net.org/). +- **[Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb):** Critical for identifying rotated objects in satellite imagery. YOLOv8, having been available longer, boasts an enormous repository of community tutorials and heavily tested enterprise deployments. If you are integrating with legacy pipelines that strictly expect YOLOv8 tensor shapes, it remains a highly dependable choice. However, for new projects prioritizing maximum efficiency—such as deploying on embedded edge devices like a Raspberry Pi—YOLO11 is the clear operational winner due to its superior speed-to-parameter ratio. @@ -115,7 +115,7 @@ Choosing between YOLOv8 and YOLO11 depends on your specific project requirements YOLOv8 is a strong choice for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. @@ -123,17 +123,17 @@ YOLOv8 is a strong choice for: YOLO11 is recommended for: -- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/) where reliability and active maintenance are paramount. -- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [OBB](https://docs.ultralytics.com/tasks/obb/) within a single unified framework. -- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python/). +- **Production Edge Deployment:** Commercial applications on devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson) where reliability and active maintenance are paramount. +- **Multi-Task Vision Applications:** Projects requiring [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [OBB](https://docs.ultralytics.com/tasks/obb) within a single unified framework. +- **Rapid Prototyping and Deployment:** Teams that need to move quickly from data collection to production using the streamlined [Ultralytics Python API](https://docs.ultralytics.com/usage/python). ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Cutting Edge: The YOLO26 Advantage diff --git a/docs/en/compare/yolov8-vs-yolo26.md b/docs/en/compare/yolov8-vs-yolo26.md index 8de021e6e35..a7ac082204a 100644 --- a/docs/en/compare/yolov8-vs-yolo26.md +++ b/docs/en/compare/yolov8-vs-yolo26.md @@ -25,9 +25,9 @@ Released in early 2023, YOLOv8 introduced a major overhaul to the YOLO framework - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2023-01-10 - **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8/) +- **Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8) -YOLOv8 quickly became the industry standard due to its excellent performance balance and deep integration into the [Ultralytics ecosystem](https://docs.ultralytics.com/). It natively supports [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [image classification](https://docs.ultralytics.com/tasks/classify/). However, it relies on standard Non-Maximum Suppression (NMS) for post-processing, which can introduce latency bottlenecks in highly constrained edge environments. +YOLOv8 quickly became the industry standard due to its excellent performance balance and deep integration into the [Ultralytics ecosystem](https://docs.ultralytics.com/). It natively supports [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [image classification](https://docs.ultralytics.com/tasks/classify). However, it relies on standard Non-Maximum Suppression (NMS) for post-processing, which can introduce latency bottlenecks in highly constrained edge environments. [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -39,9 +39,9 @@ Released in January 2026, YOLO26 takes the foundation built by its predecessors - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2026-01-14 - **GitHub:** [Ultralytics Repository](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26/) +- **Docs:** [YOLO26 Documentation](https://docs.ultralytics.com/models/yolo26) -YOLO26 introduces several paradigm-shifting technical improvements. Most notably, it features an **End-to-End NMS-Free Design**. Pioneered initially by [YOLOv10](https://docs.ultralytics.com/models/yolov10/), this architecture eliminates the need for NMS post-processing, significantly simplifying export pipelines and reducing latency variance. Furthermore, the removal of Distribution Focal Loss (DFL) streamlines the detection head, making it incredibly friendly for deployment on edge AI hardware. +YOLO26 introduces several paradigm-shifting technical improvements. Most notably, it features an **End-to-End NMS-Free Design**. Pioneered initially by [YOLOv10](https://docs.ultralytics.com/models/yolov10), this architecture eliminates the need for NMS post-processing, significantly simplifying export pipelines and reducing latency variance. Furthermore, the removal of Distribution Focal Loss (DFL) streamlines the detection head, making it incredibly friendly for deployment on edge AI hardware. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -55,15 +55,15 @@ YOLO26 brings several under-the-hood advancements that drastically improve upon ### Optimized Training with MuSGD -Training efficiency is a hallmark of Ultralytics models, which typically boast much lower memory requirements compared to bulky transformer-based architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). YOLO26 enhances this further with the introduction of the **MuSGD Optimizer**. Inspired by Large Language Model (LLM) training techniques (specifically Moonshot AI's Kimi K2), this hybrid of Stochastic Gradient Descent (SGD) and Muon ensures faster convergence and highly stable training dynamics across complex datasets. +Training efficiency is a hallmark of Ultralytics models, which typically boast much lower memory requirements compared to bulky transformer-based architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). YOLO26 enhances this further with the introduction of the **MuSGD Optimizer**. Inspired by Large Language Model (LLM) training techniques (specifically Moonshot AI's Kimi K2), this hybrid of Stochastic Gradient Descent (SGD) and Muon ensures faster convergence and highly stable training dynamics across complex datasets. ### Advanced Loss Functions -For tasks requiring high precision, such as [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensors, YOLO26 introduces **ProgLoss + STAL**. These improved loss functions provide notable enhancements in small-object recognition. Additionally, YOLO26 brings task-specific improvements across the board: a multi-scale proto for superior mask generation in segmentation, Residual Log-Likelihood Estimation (RLE) for finer pose estimation, and specialized angle loss to resolve boundary issues in [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. +For tasks requiring high precision, such as [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensors, YOLO26 introduces **ProgLoss + STAL**. These improved loss functions provide notable enhancements in small-object recognition. Additionally, YOLO26 brings task-specific improvements across the board: a multi-scale proto for superior mask generation in segmentation, Residual Log-Likelihood Estimation (RLE) for finer pose estimation, and specialized angle loss to resolve boundary issues in [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. ## Performance Analysis and Comparison -The following table highlights the performance differences between the two models using the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). Best performing values in each size category are highlighted in **bold**. +The following table highlights the performance differences between the two models using the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). Best performing values in each size category are highlighted in **bold**. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -86,13 +86,13 @@ The following table highlights the performance differences between the two model The data reveals a generational leap. YOLO26 significantly outperforms YOLOv8 across all metrics. The YOLO26 Nano (YOLO26n) model achieves a remarkable 40.9 mAP, substantially higher than YOLOv8n's 37.3, while utilizing fewer parameters and FLOPs. -One of the most striking improvements is the CPU inference speed. Because of its optimized architecture and the removal of DFL, YOLO26 delivers **up to 43% faster CPU inference** via [ONNX](https://docs.ultralytics.com/integrations/onnx/). This makes YOLO26 unparalleled for [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) and other low-resource edge devices. While GPU speeds using [TensorRT](https://developer.nvidia.com/tensorrt) are competitive in both models, the overall parameter efficiency of YOLO26 translates to lower memory footprints during both training and inference. +One of the most striking improvements is the CPU inference speed. Because of its optimized architecture and the removal of DFL, YOLO26 delivers **up to 43% faster CPU inference** via [ONNX](https://docs.ultralytics.com/integrations/onnx). This makes YOLO26 unparalleled for [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) and other low-resource edge devices. While GPU speeds using [TensorRT](https://developer.nvidia.com/tensorrt) are competitive in both models, the overall parameter efficiency of YOLO26 translates to lower memory footprints during both training and inference. ## Ease of Use and Ecosystem Both models benefit immensely from the well-maintained [Ultralytics ecosystem](https://www.ultralytics.com/). Developers praise the ease of use provided by the unified API, which allows switching between YOLOv8 and YOLO26 by simply changing the model name string. -Whether you are performing [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/), conducting [experiment tracking](https://docs.ultralytics.com/integrations/weights-biases/), or exploring new [datasets](https://docs.ultralytics.com/datasets/), the Ultralytics documentation provides extensive resources. Furthermore, the [Ultralytics Platform](https://platform.ultralytics.com/) offers a streamlined way to annotate, train, and deploy these models seamlessly into the cloud or locally. +Whether you are performing [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning), conducting [experiment tracking](https://docs.ultralytics.com/integrations/weights-biases), or exploring new [datasets](https://docs.ultralytics.com/datasets), the Ultralytics documentation provides extensive resources. Furthermore, the [Ultralytics Platform](https://platform.ultralytics.com/) offers a streamlined way to annotate, train, and deploy these models seamlessly into the cloud or locally. ### Code Example @@ -126,7 +126,7 @@ export_path = model.export(format="onnx") !!! info "Deployment Simplicity" - Exporting YOLO26 to formats like [CoreML](https://docs.ultralytics.com/integrations/coreml/) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) is significantly smoother than older models due to its NMS-free architecture, which removes complex custom operations from the exported graph. + Exporting YOLO26 to formats like [CoreML](https://docs.ultralytics.com/integrations/coreml) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino) is significantly smoother than older models due to its NMS-free architecture, which removes complex custom operations from the exported graph. ## Ideal Use Cases diff --git a/docs/en/compare/yolov8-vs-yolov10.md b/docs/en/compare/yolov8-vs-yolov10.md index da9f0e2a6c8..b086ee58885 100644 --- a/docs/en/compare/yolov8-vs-yolov10.md +++ b/docs/en/compare/yolov8-vs-yolov10.md @@ -6,7 +6,7 @@ keywords: YOLOv8 vs YOLOv10, YOLOv8 comparison, YOLOv10 performance, YOLO models # YOLOv8 vs YOLOv10: A Comprehensive Technical Comparison -The evolution of real-time [object detection](https://docs.ultralytics.com/tasks/detect/) has been moving at an unprecedented pace. As developers and researchers look to integrate the most efficient and accurate computer vision models into their pipelines, comparing leading architectures becomes essential. In this deep dive, we compare Ultralytics YOLOv8 and YOLOv10, examining their architectural differences, performance metrics, and ideal deployment scenarios to help you make an informed decision for your next AI project. +The evolution of real-time [object detection](https://docs.ultralytics.com/tasks/detect) has been moving at an unprecedented pace. As developers and researchers look to integrate the most efficient and accurate computer vision models into their pipelines, comparing leading architectures becomes essential. In this deep dive, we compare Ultralytics YOLOv8 and YOLOv10, examining their architectural differences, performance metrics, and ideal deployment scenarios to help you make an informed decision for your next AI project. @@ -23,13 +23,13 @@ Introduced as a major leap forward in the YOLO lineage, YOLOv8 established a new - Organization: [Ultralytics](https://www.ultralytics.com/) - Date: 2023-01-10 - GitHub: [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- Docs: [https://docs.ultralytics.com/models/yolov8/](https://docs.ultralytics.com/models/yolov8/) +- Docs: [https://docs.ultralytics.com/models/yolov8/](https://docs.ultralytics.com/models/yolov8) ### Architecture and Strengths YOLOv8 introduced an anchor-free detection head and a revamped CSPDarknet backbone, significantly improving both accuracy and [inference latency](https://www.ultralytics.com/glossary/inference-latency). By removing anchor boxes, the model reduces the number of box predictions, which accelerates Non-Maximum Suppression (NMS) during post-processing. -One of the standout advantages of choosing YOLOv8 is its massive versatility. While many models focus strictly on object detection, YOLOv8 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). This makes it a powerhouse for complex, multi-stage pipelines where different types of visual understanding are required simultaneously. Furthermore, its memory requirements during training are heavily optimized compared to transformer-based architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/), allowing researchers to train large models on standard consumer GPUs. +One of the standout advantages of choosing YOLOv8 is its massive versatility. While many models focus strictly on object detection, YOLOv8 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb). This makes it a powerhouse for complex, multi-stage pipelines where different types of visual understanding are required simultaneously. Furthermore, its memory requirements during training are heavily optimized compared to transformer-based architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr), allowing researchers to train large models on standard consumer GPUs. [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -44,7 +44,7 @@ Developed by researchers at Tsinghua University, YOLOv10 aimed to tackle one of - Date: 2024-05-23 - Arxiv: [https://arxiv.org/abs/2405.14458](https://arxiv.org/abs/2405.14458) - GitHub: [https://github.com/THU-MIG/yolov10](https://github.com/THU-MIG/yolov10) -- Docs: [https://docs.ultralytics.com/models/yolov10/](https://docs.ultralytics.com/models/yolov10/) +- Docs: [https://docs.ultralytics.com/models/yolov10/](https://docs.ultralytics.com/models/yolov10) ### Architecture and Strengths @@ -54,13 +54,13 @@ Additionally, YOLOv10 incorporates a holistic efficiency-accuracy driven model d !!! tip "End-to-End Detection" - The removal of NMS in YOLOv10 greatly simplifies the export process to frameworks like [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), as the entire model can be compiled as a single graph without custom post-processing layers. + The removal of NMS in YOLOv10 greatly simplifies the export process to frameworks like [OpenVINO](https://docs.ultralytics.com/integrations/openvino) and [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), as the entire model can be compiled as a single graph without custom post-processing layers. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## Performance and Metrics Comparison -When comparing these two architectures, it is crucial to look at the trade-offs between parameter count, FLOPs, and accuracy. Below is the exact comparison of their performance metrics on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). +When comparing these two architectures, it is crucial to look at the trade-offs between parameter count, FLOPs, and accuracy. Below is the exact comparison of their performance metrics on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | -------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -83,7 +83,7 @@ While YOLOv10 achieves slightly higher mAP with fewer parameters in some scales, The true differentiator for modern ML workflows is often the ecosystem surrounding the architecture. Choosing an Ultralytics model like YOLOv8 provides unparalleled **ease of use** and a seamless developer experience. -With a highly intuitive [Python SDK](https://docs.ultralytics.com/usage/python/), developers can handle data annotation, training, and deployment with minimal friction. The Ultralytics ecosystem is exceptionally well-maintained, offering frequent updates, comprehensive documentation on [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/), and robust community support on platforms like [Discord](https://discord.com/invite/ultralytics) and GitHub. +With a highly intuitive [Python SDK](https://docs.ultralytics.com/usage/python), developers can handle data annotation, training, and deployment with minimal friction. The Ultralytics ecosystem is exceptionally well-maintained, offering frequent updates, comprehensive documentation on [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning), and robust community support on platforms like [Discord](https://discord.com/invite/ultralytics) and GitHub. ### Code Example: Simplified Training @@ -120,7 +120,7 @@ Choosing between YOLOv8 and YOLOv10 depends on your specific project requirement YOLOv8 is a strong choice for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. @@ -134,24 +134,24 @@ YOLOv10 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Future: Stepping Up to YOLO26 While YOLOv8 is a fantastic all-rounder and YOLOv10 provides great academic insights into NMS-free architectures, the cutting edge of computer vision has moved forward. For the ultimate balance of speed, accuracy, and deployment simplicity, we strongly recommend migrating to **YOLO26**. -Released in early 2026, [YOLO26](https://docs.ultralytics.com/models/yolo26/) represents the absolute pinnacle of the YOLO family. It seamlessly merges the best features of its predecessors while introducing groundbreaking new technologies: +Released in early 2026, [YOLO26](https://docs.ultralytics.com/models/yolo26) represents the absolute pinnacle of the YOLO family. It seamlessly merges the best features of its predecessors while introducing groundbreaking new technologies: - **End-to-End NMS-Free Design:** Adopting the breakthrough pioneered by YOLOv10, YOLO26 natively eliminates NMS for faster, simpler deployment. -- **DFL Removal:** The removal of Distribution Focal Loss makes exporting the model to [CoreML](https://docs.ultralytics.com/integrations/coreml/) and edge devices significantly smoother. +- **DFL Removal:** The removal of Distribution Focal Loss makes exporting the model to [CoreML](https://docs.ultralytics.com/integrations/coreml) and edge devices significantly smoother. - **MuSGD Optimizer:** Inspired by Large Language Model (LLM) training paradigms, this hybrid optimizer guarantees faster convergence and unmatched training stability. -- **CPU Inference Dominance:** YOLO26 delivers up to 43% faster CPU inference compared to previous generations, making it a game-changer for [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) and IoT applications. +- **CPU Inference Dominance:** YOLO26 delivers up to 43% faster CPU inference compared to previous generations, making it a game-changer for [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) and IoT applications. - **ProgLoss + STAL:** These advanced loss functions provide notable improvements in small-object recognition, which is critical for aerial imagery and robotics. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } -If you are currently evaluating models, you might also be interested in [YOLO11](https://docs.ultralytics.com/models/yolo11/), the direct predecessor to YOLO26, which remains a rock-solid, production-ready framework widely used in enterprise solutions today. However, for maximum future-proofing and performance, exploring the advanced capabilities of the [Ultralytics Platform](https://platform.ultralytics.com/) with YOLO26 is the best path forward for your vision AI strategy. +If you are currently evaluating models, you might also be interested in [YOLO11](https://docs.ultralytics.com/models/yolo11), the direct predecessor to YOLO26, which remains a rock-solid, production-ready framework widely used in enterprise solutions today. However, for maximum future-proofing and performance, exploring the advanced capabilities of the [Ultralytics Platform](https://platform.ultralytics.com/) with YOLO26 is the best path forward for your vision AI strategy. diff --git a/docs/en/compare/yolov8-vs-yolov5.md b/docs/en/compare/yolov8-vs-yolov5.md index 17e50718751..10f2944b98c 100644 --- a/docs/en/compare/yolov8-vs-yolov5.md +++ b/docs/en/compare/yolov8-vs-yolov5.md @@ -6,7 +6,7 @@ keywords: YOLOv8, YOLOv5, object detection, YOLO comparison, computer vision, mo # YOLOv8 vs. YOLOv5: A Comprehensive Technical Comparison -Choosing the right computer vision architecture is a critical step in building robust machine learning pipelines. In this detailed technical comparison, we explore the differences between two of the most popular models in the vision AI ecosystem: **YOLOv8** and **YOLOv5**. Both models were developed by Ultralytics and have significantly shaped the landscape of real-time [object detection](https://docs.ultralytics.com/tasks/detect/), setting industry standards for speed, accuracy, and ease of use. +Choosing the right computer vision architecture is a critical step in building robust machine learning pipelines. In this detailed technical comparison, we explore the differences between two of the most popular models in the vision AI ecosystem: **YOLOv8** and **YOLOv5**. Both models were developed by Ultralytics and have significantly shaped the landscape of real-time [object detection](https://docs.ultralytics.com/tasks/detect), setting industry standards for speed, accuracy, and ease of use. Whether you are deploying to edge devices or scaling cloud inference, understanding the architectural shifts, performance metrics, and training methodologies of these models will help you make an informed decision for your [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) projects. @@ -17,13 +17,13 @@ Whether you are deploying to edge devices or scaling cloud inference, understand ## Ultralytics YOLOv8: The Versatile Standard -Released in early 2023, YOLOv8 represented a major architectural shift from its predecessors. It was designed from the ground up to serve as a unified framework capable of handling multiple vision tasks natively, including [instance segmentation](https://docs.ultralytics.com/tasks/segment/), image classification, and [pose estimation](https://docs.ultralytics.com/tasks/pose/). +Released in early 2023, YOLOv8 represented a major architectural shift from its predecessors. It was designed from the ground up to serve as a unified framework capable of handling multiple vision tasks natively, including [instance segmentation](https://docs.ultralytics.com/tasks/segment), image classification, and [pose estimation](https://docs.ultralytics.com/tasks/pose). - **Authors:** Glenn Jocher, Ayush Chaurasia, and Jing Qiu - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2023-01-10 - **GitHub:** [ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8/) +- **Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8) ### Architecture and Methodologies @@ -33,7 +33,7 @@ The architecture features a **C2f module** (Cross-Stage Partial bottleneck with !!! tip "Memory Efficiency" - Ultralytics YOLO models, including YOLOv8, are optimized for lower CUDA memory usage during training compared to many Transformer-based alternatives like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). This allows developers to use larger batch sizes on standard consumer GPUs like the NVIDIA RTX series. + Ultralytics YOLO models, including YOLOv8, are optimized for lower CUDA memory usage during training compared to many Transformer-based alternatives like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). This allows developers to use larger batch sizes on standard consumer GPUs like the NVIDIA RTX series. ### Strengths and Weaknesses @@ -57,7 +57,7 @@ Introduced in 2020, YOLOv5 brought YOLO to the [PyTorch](https://pytorch.org/) e - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2020-06-26 - **GitHub:** [ultralytics/yolov5](https://github.com/ultralytics/yolov5) -- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5/) +- **Docs:** [YOLOv5 Documentation](https://docs.ultralytics.com/models/yolov5) ### Architecture and Methodologies @@ -82,7 +82,7 @@ YOLOv5 incorporates the **C3 module**, which efficiently extracts features while ## Performance Comparison -When evaluating these models, achieving a favorable trade-off between speed and accuracy is paramount. The table below outlines the performance metrics of both architectures evaluated on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). CPU speeds were measured using [ONNX](https://onnx.ai/), while GPU speeds were tested using [TensorRT](https://developer.nvidia.com/tensorrt). +When evaluating these models, achieving a favorable trade-off between speed and accuracy is paramount. The table below outlines the performance metrics of both architectures evaluated on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). CPU speeds were measured using [ONNX](https://onnx.ai/), while GPU speeds were tested using [TensorRT](https://developer.nvidia.com/tensorrt). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -104,7 +104,7 @@ While YOLOv5 retains a slight edge in parameter count and absolute raw speed for A defining characteristic of modern Ultralytics models is the well-maintained ecosystem surrounding them. The transition from YOLOv5 to YOLOv8 brought the introduction of the unified `ultralytics` pip package, creating a highly streamlined user experience. -Developers can seamlessly handle [model training](https://docs.ultralytics.com/modes/train/), validation, prediction, and export with just a few lines of Python code, bypassing the complex boilerplate scripts historically required in deep learning projects. +Developers can seamlessly handle [model training](https://docs.ultralytics.com/modes/train), validation, prediction, and export with just a few lines of Python code, bypassing the complex boilerplate scripts historically required in deep learning projects. ```python from ultralytics import YOLO @@ -124,10 +124,10 @@ Furthermore, integration with tools like [Ultralytics Platform](https://platform ## Ideal Use Cases **When to choose YOLOv5:** -If you are maintaining legacy systems, running inference on severely constrained CPUs like a [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/), or working on a project where saving every fraction of a megabyte in model size is critical, YOLOv5 remains a reliable workhorse. +If you are maintaining legacy systems, running inference on severely constrained CPUs like a [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi), or working on a project where saving every fraction of a megabyte in model size is critical, YOLOv5 remains a reliable workhorse. **When to choose YOLOv8:** -For virtually all new projects starting today, YOLOv8 is highly recommended over YOLOv5. Its advanced architecture handles complex tracking, [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), and segmentation effortlessly. It is ideal for modern applications spanning from autonomous robotics to medical image analysis and smart city infrastructure. +For virtually all new projects starting today, YOLOv8 is highly recommended over YOLOv5. Its advanced architecture handles complex tracking, [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb), and segmentation effortlessly. It is ideal for modern applications spanning from autonomous robotics to medical image analysis and smart city infrastructure. !!! info "Looking for the Latest State-of-the-Art?" @@ -139,4 +139,4 @@ For virtually all new projects starting today, YOLOv8 is highly recommended over * **DFL Removal:** Distribution Focal Loss has been removed for simplified export and enhanced edge device compatibility. * **ProgLoss + STAL:** Advanced loss functions that drive notable improvements in small-object recognition, which is critical for aerial imagery and IoT. -By leveraging the comprehensive documentation and tools provided by Ultralytics, you can easily deploy YOLOv8, or explore the cutting-edge YOLO26, to solve complex visual challenges with unprecedented speed and accuracy. For further learning, consider exploring our guides on [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/) and [model deployment practices](https://docs.ultralytics.com/guides/model-deployment-practices/). +By leveraging the comprehensive documentation and tools provided by Ultralytics, you can easily deploy YOLOv8, or explore the cutting-edge YOLO26, to solve complex visual challenges with unprecedented speed and accuracy. For further learning, consider exploring our guides on [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning) and [model deployment practices](https://docs.ultralytics.com/guides/model-deployment-practices). diff --git a/docs/en/compare/yolov8-vs-yolov6.md b/docs/en/compare/yolov8-vs-yolov6.md index 83b6c1d10d3..47cb55dd06c 100644 --- a/docs/en/compare/yolov8-vs-yolov6.md +++ b/docs/en/compare/yolov8-vs-yolov6.md @@ -8,7 +8,7 @@ keywords: YOLOv8, YOLOv6-3.0, object detection, machine learning, computer visio The landscape of real-time computer vision is constantly evolving, driven by the demand for faster, more accurate, and more versatile models. Two of the most prominent architectures that emerged in early 2023 are [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) and YOLOv6-3.0 by Meituan. Both models push the boundaries of state-of-the-art performance, but they cater to slightly different development philosophies and deployment scenarios. -This comprehensive guide provides an in-depth analysis of their architectures, performance metrics, and ideal use cases, helping machine learning engineers and researchers choose the right tool for their next [object detection](https://docs.ultralytics.com/tasks/detect/) project. +This comprehensive guide provides an in-depth analysis of their architectures, performance metrics, and ideal use cases, helping machine learning engineers and researchers choose the right tool for their next [object detection](https://docs.ultralytics.com/tasks/detect) project. @@ -27,7 +27,7 @@ The Ultralytics YOLOv8 architecture represents a unified, multi-task framework d - Organization: [Ultralytics](https://www.ultralytics.com/) - Date: 2023-01-10 - GitHub: [https://github.com/ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- Docs: [https://docs.ultralytics.com/models/yolov8/](https://docs.ultralytics.com/models/yolov8/) +- Docs: [https://docs.ultralytics.com/models/yolov8/](https://docs.ultralytics.com/models/yolov8) [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -40,9 +40,9 @@ Originally introduced for industrial applications at Meituan, YOLOv6 received a - Date: 2023-01-13 - Arxiv: [https://arxiv.org/abs/2301.05586](https://arxiv.org/abs/2301.05586) - GitHub: [https://github.com/meituan/YOLOv6](https://github.com/meituan/YOLOv6) -- Docs: [https://docs.ultralytics.com/models/yolov6/](https://docs.ultralytics.com/models/yolov6/) +- Docs: [https://docs.ultralytics.com/models/yolov6/](https://docs.ultralytics.com/models/yolov6) -[Learn more about YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6){ .md-button } !!! tip "Streamlined Management" @@ -54,7 +54,7 @@ Originally introduced for industrial applications at Meituan, YOLOv6 received a YOLOv8 introduced a highly refined, anchor-free detection head. By removing predefined anchor boxes, the model generalizes better across diverse datasets and reduces the number of post-processing heuristics. Furthermore, YOLOv8 offers an unmatched **Performance Balance**, consistently achieving a favorable trade-off between speed and accuracy suitable for diverse real-world deployment scenarios—from cloud servers to resource-constrained edge devices. -A major advantage of YOLOv8 is its **Memory requirements**. During training, Ultralytics models exhibit significantly lower CUDA memory usage compared to heavy transformer-based alternatives like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). This allows developers to utilize larger batch sizes on standard consumer GPUs, resulting in excellent **Training Efficiency**. +A major advantage of YOLOv8 is its **Memory requirements**. During training, Ultralytics models exhibit significantly lower CUDA memory usage compared to heavy transformer-based alternatives like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). This allows developers to utilize larger batch sizes on standard consumer GPUs, resulting in excellent **Training Efficiency**. ### The YOLOv6-3.0 Architecture @@ -83,7 +83,7 @@ While YOLOv6-3.0 boasts slight speed advantages on specific TensorRT benchmarks, The starkest contrast between the two models lies in their ecosystem support. -YOLOv6 is primarily a bounding-box detection engine. In contrast, YOLOv8 is celebrated for its **Versatility**. Within a single unified framework, YOLOv8 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. +YOLOv6 is primarily a bounding-box detection engine. In contrast, YOLOv8 is celebrated for its **Versatility**. Within a single unified framework, YOLOv8 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. Furthermore, the **Ease of Use** of the Ultralytics ecosystem is unparalleled. With a simple Python API, researchers can initiate training, validate results, and export models to numerous formats without writing complex boilerplate code. The **Well-Maintained Ecosystem** ensures active development, frequent updates, and seamless integrations with popular experiment tracking tools. @@ -118,7 +118,7 @@ Choosing between YOLOv8 and YOLOv6 depends on your specific project requirements YOLOv8 is a strong choice for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. @@ -132,11 +132,11 @@ YOLOv6 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Looking Forward: Upgrading to YOLO26 @@ -150,7 +150,7 @@ Training stability and convergence speed have also seen massive upgrades thanks !!! note "Other Models to Consider" - Depending on your specific constraints, you may also be interested in exploring [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) for highly balanced legacy workflows or [YOLO-World](https://docs.ultralytics.com/models/yolo-world/) for zero-shot, open-vocabulary detection tasks without the need for extensive retraining. + Depending on your specific constraints, you may also be interested in exploring [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) for highly balanced legacy workflows or [YOLO-World](https://docs.ultralytics.com/models/yolo-world) for zero-shot, open-vocabulary detection tasks without the need for extensive retraining. ## Conclusion diff --git a/docs/en/compare/yolov8-vs-yolov7.md b/docs/en/compare/yolov8-vs-yolov7.md index 5d0c721ffe1..6459caf23b7 100644 --- a/docs/en/compare/yolov8-vs-yolov7.md +++ b/docs/en/compare/yolov8-vs-yolov7.md @@ -23,7 +23,7 @@ Released by Ultralytics in early 2023, YOLOv8 represents a major architectural s - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2023-01-10 - **GitHub:** [ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8/) +- **Docs:** [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8) ### Architectural Innovations @@ -53,7 +53,7 @@ YOLOv7 employs an **Extended Efficient Layer Aggregation Network (E-ELAN)**, whi While YOLOv7 achieves excellent performance on standard academic benchmarks like the [MS COCO dataset](https://cocodataset.org/), its architecture is heavily optimized for server-grade accelerators. Exporting and deploying these models to edge devices can sometimes require more manual configuration compared to more modern, streamlined frameworks. -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ## Detailed Performance Comparison @@ -78,7 +78,7 @@ While YOLOv7 provides strong raw detection metrics, **Ultralytics YOLOv8** outsh ### Unmatched Versatility -YOLOv7 is primarily a detection model, with experimental branches for other tasks. In contrast, YOLOv8 natively supports [Object Detection](https://docs.ultralytics.com/tasks/detect/), [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). This unified approach means a team can learn one API and deploy it across entirely different project requirements. +YOLOv7 is primarily a detection model, with experimental branches for other tasks. In contrast, YOLOv8 natively supports [Object Detection](https://docs.ultralytics.com/tasks/detect), [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Image Classification](https://docs.ultralytics.com/tasks/classify), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). This unified approach means a team can learn one API and deploy it across entirely different project requirements. ### Streamlined Deployment and Integrations diff --git a/docs/en/compare/yolov8-vs-yolov9.md b/docs/en/compare/yolov8-vs-yolov9.md index 4793b19c050..762b3980f97 100644 --- a/docs/en/compare/yolov8-vs-yolov9.md +++ b/docs/en/compare/yolov8-vs-yolov9.md @@ -6,7 +6,7 @@ keywords: YOLOv8, YOLOv9, object detection, model comparison, Ultralytics, perfo # YOLOv8 vs. YOLOv9: A Comprehensive Technical Comparison of Real-Time Object Detectors -The evolution of real-time object detection has been characterized by a constant push for better accuracy, lower latency, and improved hardware utilization. Two major milestones in this journey are [Ultralytics YOLOv8](https://docs.ultralytics.com/models/yolov8/) and [YOLOv9](https://docs.ultralytics.com/models/yolov9/). While both models represent state-of-the-art capabilities in [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv), they cater to different deployment needs, architectural philosophies, and developer ecosystems. +The evolution of real-time object detection has been characterized by a constant push for better accuracy, lower latency, and improved hardware utilization. Two major milestones in this journey are [Ultralytics YOLOv8](https://docs.ultralytics.com/models/yolov8) and [YOLOv9](https://docs.ultralytics.com/models/yolov9). While both models represent state-of-the-art capabilities in [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv), they cater to different deployment needs, architectural philosophies, and developer ecosystems. This comprehensive guide breaks down the technical differences, architectural innovations, and practical deployment considerations to help you choose the right model for your next artificial intelligence project. @@ -27,7 +27,7 @@ Released by the team at [Ultralytics](https://www.ultralytics.com/about), YOLOv8 - **Organization:** [Ultralytics](https://www.ultralytics.com/) - **Date:** 2023-01-10 - **GitHub:** [ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -- **Documentation:** [YOLOv8 Docs](https://docs.ultralytics.com/models/yolov8/) +- **Documentation:** [YOLOv8 Docs](https://docs.ultralytics.com/models/yolov8) [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -41,7 +41,7 @@ Developed independently by researchers at Academia Sinica, YOLOv9 focuses heavil - **Arxiv:** [2402.13616](https://arxiv.org/abs/2402.13616) - **GitHub:** [WongKinYiu/yolov9](https://github.com/WongKinYiu/yolov9) -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } !!! tip "Enterprise Deployment" @@ -55,7 +55,7 @@ The architectural choices in deep learning dictate how efficiently a model learn YOLOv8 introduced the **C2f module** (Cross-Stage Partial bottleneck with two convolutions), which replaced the older C3 module. This change improves gradient flow and allows the network to learn richer feature representations without heavily taxing [GPU memory](https://www.ultralytics.com/glossary/gpu-graphics-processing-unit). -Furthermore, YOLOv8 utilizes an **anchor-free** design with a **decoupled head**. By processing objectness, classification, and regression through separate pathways, the model converges faster during training and generalizes better to diverse [custom datasets](https://docs.ultralytics.com/datasets/). +Furthermore, YOLOv8 utilizes an **anchor-free** design with a **decoupled head**. By processing objectness, classification, and regression through separate pathways, the model converges faster during training and generalizes better to diverse [custom datasets](https://docs.ultralytics.com/datasets). ### YOLOv9 Architecture: PGI and GELAN @@ -85,7 +85,7 @@ _Note: Best values in each column are highlighted in **bold**._ ### Analyzing the Trade-offs -YOLOv9 achieves slightly higher peak accuracy (mAP), particularly with its larger `e` variant. However, this comes at a cost. Ultralytics YOLOv8 maintains a significant advantage in **inference speed**, particularly when compiled to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) or [ONNX](https://docs.ultralytics.com/integrations/onnx/). For applications requiring high frames-per-second (FPS) on constrained edge hardware (like a [Raspberry Pi](https://www.raspberrypi.org/) or older mobile chips), YOLOv8's `n` and `s` variants offer a far more practical performance balance. +YOLOv9 achieves slightly higher peak accuracy (mAP), particularly with its larger `e` variant. However, this comes at a cost. Ultralytics YOLOv8 maintains a significant advantage in **inference speed**, particularly when compiled to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) or [ONNX](https://docs.ultralytics.com/integrations/onnx). For applications requiring high frames-per-second (FPS) on constrained edge hardware (like a [Raspberry Pi](https://www.raspberrypi.org/) or older mobile chips), YOLOv8's `n` and `s` variants offer a far more practical performance balance. ## Training Efficiency and Ecosystem Integration @@ -114,7 +114,7 @@ This single API dramatically reduces the time from prototype to production. Furt ### Task Versatility -While YOLOv9 is an excellent bounding box detector, real-world vision AI often requires more. YOLOv8 is a versatile powerhouse natively supporting [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). Using a single framework for multiple tasks drastically reduces software bloat and maintenance overhead. +While YOLOv9 is an excellent bounding box detector, real-world vision AI often requires more. YOLOv8 is a versatile powerhouse natively supporting [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Image Classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). Using a single framework for multiple tasks drastically reduces software bloat and maintenance overhead. !!! note "Looking Forward" @@ -126,7 +126,7 @@ How do these models fare in production? ### Autonomous Drones and Robotics -For robotics requiring rapid obstacle avoidance, **YOLOv8** is the preferred choice. The ultra-low latency of `YOLOv8n` ensures that autonomous systems react to their environments in real-time, preventing collisions. The native export capabilities to [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) and CoreML make it trivial to deploy on the low-power chips typical of commercial drones. +For robotics requiring rapid obstacle avoidance, **YOLOv8** is the preferred choice. The ultra-low latency of `YOLOv8n` ensures that autonomous systems react to their environments in real-time, preventing collisions. The native export capabilities to [OpenVINO](https://docs.ultralytics.com/integrations/openvino) and CoreML make it trivial to deploy on the low-power chips typical of commercial drones. ### High-Resolution Defect Detection @@ -134,7 +134,7 @@ In specialized manufacturing settings where detecting microscopic anomalies is c ### Smart Retail and Security Analytics -For tracking customers across store aisles or managing [automated checkout systems](https://www.ultralytics.com/solutions/ai-in-retail), **YOLOv8** provides the best balance. Its ability to simultaneously run detection and [multi-object tracking](https://docs.ultralytics.com/modes/track/) using standard algorithms like BoT-SORT makes it a robust solution for multi-camera retail deployments. +For tracking customers across store aisles or managing [automated checkout systems](https://www.ultralytics.com/solutions/ai-in-retail), **YOLOv8** provides the best balance. Its ability to simultaneously run detection and [multi-object tracking](https://docs.ultralytics.com/modes/track) using standard algorithms like BoT-SORT makes it a robust solution for multi-camera retail deployments. ## Use Cases and Recommendations @@ -144,7 +144,7 @@ Choosing between YOLOv8 and YOLOv9 depends on your specific project requirements YOLOv8 is a strong choice for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. @@ -158,16 +158,16 @@ YOLOv9 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Next Evolution: YOLO26 While YOLOv8 and YOLOv9 are powerful, the AI landscape moves rapidly. For teams demanding the absolute best performance, the newly released **YOLO26** builds upon the successes of these previous generations. -YOLO26 introduces an **end-to-end NMS-free design**, which completely eliminates complex post-processing bottlenecks, making deployment simpler and latency more predictable. Driven by the new **MuSGD Optimizer** and enhanced **ProgLoss + STAL** loss functions, and with **DFL Removal** (Distribution Focal Loss removed for simplified export and better edge/low-power device compatibility), it achieves up to **43% faster CPU inference** while boosting small-object recognition. For developers pushing the limits of edge computing, evaluating [YOLO26](https://docs.ultralytics.com/models/yolo26/) is highly recommended. +YOLO26 introduces an **end-to-end NMS-free design**, which completely eliminates complex post-processing bottlenecks, making deployment simpler and latency more predictable. Driven by the new **MuSGD Optimizer** and enhanced **ProgLoss + STAL** loss functions, and with **DFL Removal** (Distribution Focal Loss removed for simplified export and better edge/low-power device compatibility), it achieves up to **43% faster CPU inference** while boosting small-object recognition. For developers pushing the limits of edge computing, evaluating [YOLO26](https://docs.ultralytics.com/models/yolo26) is highly recommended. In summary, while YOLOv9 offers fascinating architectural research and excellent peak accuracy, **Ultralytics YOLOv8** remains the most practical, well-supported, and versatile choice for the vast majority of computer vision engineers aiming to ship reliable software quickly. diff --git a/docs/en/compare/yolov8-vs-yolox.md b/docs/en/compare/yolov8-vs-yolox.md index 1721b640a0e..b4315930b9f 100644 --- a/docs/en/compare/yolov8-vs-yolox.md +++ b/docs/en/compare/yolov8-vs-yolox.md @@ -6,7 +6,7 @@ keywords: YOLOv8, YOLOX, object detection, model comparison, Ultralytics, comput # YOLOv8 vs YOLOX: Analyzing Anchor-Free Object Detection Models -The landscape of computer vision has been heavily shaped by the continuous evolution of real-time object detection architectures. Two prominent milestones in this journey are [Ultralytics YOLOv8](https://docs.ultralytics.com/models/yolov8/) and YOLOX. While both models embrace an anchor-free design paradigm to streamline bounding box predictions, they represent different eras and philosophies in deep learning research and deployment ecosystem development. +The landscape of computer vision has been heavily shaped by the continuous evolution of real-time object detection architectures. Two prominent milestones in this journey are [Ultralytics YOLOv8](https://docs.ultralytics.com/models/yolov8) and YOLOX. While both models embrace an anchor-free design paradigm to streamline bounding box predictions, they represent different eras and philosophies in deep learning research and deployment ecosystem development. @@ -21,9 +21,9 @@ Understanding the origins and design goals of each framework provides critical c ### Ultralytics YOLOv8 -Developed by Glenn Jocher, Ayush Chaurasia, and Jing Qiu at Ultralytics and released on January 10, 2023, YOLOv8 marked a significant leap in the Ultralytics ecosystem. Building upon the massive success of [YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5), YOLOv8 introduced a highly refined, state-of-the-art architecture capable of handling a diverse array of tasks natively, including [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/). +Developed by Glenn Jocher, Ayush Chaurasia, and Jing Qiu at Ultralytics and released on January 10, 2023, YOLOv8 marked a significant leap in the Ultralytics ecosystem. Building upon the massive success of [YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5), YOLOv8 introduced a highly refined, state-of-the-art architecture capable of handling a diverse array of tasks natively, including [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose). -Its primary advantage lies in the well-maintained Ultralytics ecosystem, which provides a seamless "zero-to-hero" experience with a unified Python API, extensive documentation, and native integrations with MLOps tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) and [Comet](https://docs.ultralytics.com/integrations/comet/). +Its primary advantage lies in the well-maintained Ultralytics ecosystem, which provides a seamless "zero-to-hero" experience with a unified Python API, extensive documentation, and native integrations with MLOps tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) and [Comet](https://docs.ultralytics.com/integrations/comet). [Explore YOLOv8 on the Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -96,13 +96,13 @@ metrics = model.val() model.export(format="onnx") ``` -This API standardizes workflows across detection, segmentation, and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb/) tasks, drastically reducing time-to-market for production applications. Furthermore, built-in [export functionalities](https://docs.ultralytics.com/modes/export/) allow seamless conversion to [ONNX](https://docs.ultralytics.com/integrations/onnx/), [OpenVINO](https://docs.ultralytics.com/integrations/openvino/), and CoreML without writing custom C++ operators. +This API standardizes workflows across detection, segmentation, and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb) tasks, drastically reducing time-to-market for production applications. Furthermore, built-in [export functionalities](https://docs.ultralytics.com/modes/export) allow seamless conversion to [ONNX](https://docs.ultralytics.com/integrations/onnx), [OpenVINO](https://docs.ultralytics.com/integrations/openvino), and CoreML without writing custom C++ operators. ## Ideal Use Cases Choosing between these architectures depends on your project constraints, though YOLOv8 provides a much more flexible foundation. -- **High-Speed Edge Analytics:** For real-time processing on devices like the [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson/), YOLOv8 offers an unmatched balance of speed and accuracy, easily deployable via its native TensorRT integration. +- **High-Speed Edge Analytics:** For real-time processing on devices like the [NVIDIA Jetson](https://docs.ultralytics.com/guides/nvidia-jetson), YOLOv8 offers an unmatched balance of speed and accuracy, easily deployable via its native TensorRT integration. - **Academic Research:** YOLOX remains a valuable educational tool for researchers studying the transition from anchor-based to anchor-free methodologies within PyTorch. - **Complex Multi-Task Applications:** Applications requiring simultaneous object tracking and instance segmentation will heavily favor YOLOv8, as these capabilities are built directly into the Ultralytics library. diff --git a/docs/en/compare/yolov9-vs-damo-yolo.md b/docs/en/compare/yolov9-vs-damo-yolo.md index fe1f49fa868..d75b97962b9 100644 --- a/docs/en/compare/yolov9-vs-damo-yolo.md +++ b/docs/en/compare/yolov9-vs-damo-yolo.md @@ -26,9 +26,9 @@ YOLOv9 was designed to directly address the information loss that occurs as data **Authors:** Chien-Yao Wang, Hong-Yuan Mark Liao **Organization:** Institute of Information Science, Academia Sinica, Taiwan **Date:** February 21, 2024 -**Links:** [Arxiv](https://arxiv.org/abs/2402.13616), [GitHub](https://github.com/WongKinYiu/yolov9), [Docs](https://docs.ultralytics.com/models/yolov9/) +**Links:** [Arxiv](https://arxiv.org/abs/2402.13616), [GitHub](https://github.com/WongKinYiu/yolov9), [Docs](https://docs.ultralytics.com/models/yolov9) -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } YOLOv9 introduces **Programmable Gradient Information (PGI)** and the **Generalized Efficient Layer Aggregation Network (GELAN)**. PGI ensures that vital spatial and semantic information is retained during the feed-forward process, preventing the degradation of gradients used for weight updates. GELAN complements this by maximizing parameter efficiency, allowing the model to achieve state-of-the-art [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map) with fewer FLOPs than many conventional CNNs. @@ -51,7 +51,7 @@ DAMO-YOLO relies on a MAE-NAS (Masked Autoencoders for Neural Architecture Searc ## Performance and Metrics Comparison -When selecting an [object detection](https://docs.ultralytics.com/tasks/detect/) model, balancing accuracy, speed, and computational footprint is critical. +When selecting an [object detection](https://docs.ultralytics.com/tasks/detect) model, balancing accuracy, speed, and computational footprint is critical. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -69,7 +69,7 @@ When selecting an [object detection](https://docs.ultralytics.com/tasks/detect/) ### Analysis - **Accuracy vs. Parameters:** YOLOv9 generally demonstrates a superior parameter-to-accuracy ratio. For instance, YOLOv9c achieves 53.0% mAP with 25.3M parameters, while DAMO-YOLOl achieves 50.8% mAP but requires significantly more parameters (42.1M). -- **Inference Speed:** DAMO-YOLO's architecture provides competitive TensorRT inference speeds on T4 GPUs, slightly edging out YOLOv9 in the medium tiers. However, YOLOv9's efficiency in FLOPs and parameter count translates to exceptional [GPU memory efficiency](https://docs.ultralytics.com/guides/yolo-performance-metrics/). +- **Inference Speed:** DAMO-YOLO's architecture provides competitive TensorRT inference speeds on T4 GPUs, slightly edging out YOLOv9 in the medium tiers. However, YOLOv9's efficiency in FLOPs and parameter count translates to exceptional [GPU memory efficiency](https://docs.ultralytics.com/guides/yolo-performance-metrics). - **Memory Requirements:** Ultralytics YOLO models, including YOLOv9, typically exhibit lower memory usage during both training and inference compared to complex NAS-generated models or heavy transformer architectures, making them highly accessible for deployment on constrained edge hardware. ## The Ultralytics Ecosystem Advantage @@ -78,7 +78,7 @@ While theoretical metrics are important, practical implementation heavily dictat ### Ease of Use and Training Efficiency -Training a custom YOLOv9 model requires minimal boilerplate. The [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) abstracts complex processes like [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation/), distributed training, and hardware optimization. +Training a custom YOLOv9 model requires minimal boilerplate. The [Ultralytics Python API](https://docs.ultralytics.com/usage/python) abstracts complex processes like [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation), distributed training, and hardware optimization. ```python from ultralytics import YOLO @@ -100,11 +100,11 @@ Conversely, utilizing DAMO-YOLO often requires navigating rigid configuration fi ### Versatility Across Tasks -A hallmark of Ultralytics models is their inherent versatility. Beyond standard bounding box detection, the Ultralytics framework seamlessly supports tasks such as [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. DAMO-YOLO is strictly optimized for 2D object detection, requiring significant re-engineering to adapt to other visual paradigms. +A hallmark of Ultralytics models is their inherent versatility. Beyond standard bounding box detection, the Ultralytics framework seamlessly supports tasks such as [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Image Classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. DAMO-YOLO is strictly optimized for 2D object detection, requiring significant re-engineering to adapt to other visual paradigms. !!! tip "Exporting to Edge Devices" - Ultralytics simplifies the deployment pipeline by offering one-click [model export](https://docs.ultralytics.com/modes/export/) to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), [OpenVINO](https://docs.ultralytics.com/integrations/openvino/), and CoreML, ensuring maximum performance regardless of your target hardware. + Ultralytics simplifies the deployment pipeline by offering one-click [model export](https://docs.ultralytics.com/modes/export) to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), [OpenVINO](https://docs.ultralytics.com/integrations/openvino), and CoreML, ensuring maximum performance regardless of your target hardware. ## Use Cases and Recommendations @@ -128,11 +128,11 @@ DAMO-YOLO is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Future: Moving to YOLO26 @@ -142,7 +142,7 @@ Released in 2026, YOLO26 builds upon the successes of its predecessors, offering ### Key YOLO26 Innovations -- **End-to-End NMS-Free Design:** YOLO26 eliminates Non-Maximum Suppression (NMS) post-processing entirely. This creates a streamlined deployment pipeline that is natively end-to-end, a breakthrough first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/). +- **End-to-End NMS-Free Design:** YOLO26 eliminates Non-Maximum Suppression (NMS) post-processing entirely. This creates a streamlined deployment pipeline that is natively end-to-end, a breakthrough first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10). - **DFL Removal:** Distribution Focal Loss removed for simplified export and better edge/low-power device compatibility. - **Up to 43% Faster CPU Inference:** By removing complex post-processing and optimizing core convolutions, YOLO26 is uniquely suited for edge computing scenarios lacking dedicated GPUs. - **MuSGD Optimizer:** Inspired by LLM training innovations, YOLO26 utilizes a hybrid of SGD and Muon (MuSGD) to guarantee more stable training runs and noticeably faster convergence times. diff --git a/docs/en/compare/yolov9-vs-efficientdet.md b/docs/en/compare/yolov9-vs-efficientdet.md index 8ce6a7b57df..e85b40bcbfa 100644 --- a/docs/en/compare/yolov9-vs-efficientdet.md +++ b/docs/en/compare/yolov9-vs-efficientdet.md @@ -6,7 +6,7 @@ keywords: YOLOv9,EfficientDet,object detection,model comparison,YOLO,EfficientDe # YOLOv9 vs. EfficientDet: A Comprehensive Technical Comparison of Object Detection Architectures -The field of computer vision has witnessed a rapid evolution in real-time [object detection](https://docs.ultralytics.com/tasks/detect/), with researchers continuously pushing the boundaries of accuracy and efficiency. When building robust vision systems, selecting the optimal architecture is a critical decision. Two highly discussed models in this space are **YOLOv9**, an advanced iteration of the YOLO lineage focusing on gradient information, and **EfficientDet**, a scalable framework developed by Google. +The field of computer vision has witnessed a rapid evolution in real-time [object detection](https://docs.ultralytics.com/tasks/detect), with researchers continuously pushing the boundaries of accuracy and efficiency. When building robust vision systems, selecting the optimal architecture is a critical decision. Two highly discussed models in this space are **YOLOv9**, an advanced iteration of the YOLO lineage focusing on gradient information, and **EfficientDet**, a scalable framework developed by Google. This guide provides an in-depth technical analysis comparing these two architectures, examining their underlying mechanics, performance metrics, and ideal deployment scenarios to help you make an informed decision for your next AI project. @@ -30,7 +30,7 @@ Developed to tackle the deep learning "information bottleneck," YOLOv9 introduce YOLOv9 introduces **Programmable Gradient Information (PGI)**, an auxiliary supervision framework that guarantees gradient information is reliably preserved across deep layers. This is coupled with the **Generalized Efficient Layer Aggregation Network (GELAN)**, which optimizes parameter efficiency by combining the strengths of CSPNet and ELAN. This allows YOLOv9 to achieve high [accuracy](https://www.ultralytics.com/glossary/accuracy) while maintaining a lightweight footprint suitable for real-time edge processing. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ### EfficientDet: Compound Scaling and BiFPN @@ -47,7 +47,7 @@ EfficientDet relies on an EfficientNet backbone combined with a **Bidirectional !!! tip "Choosing the Right Framework" - While theoretical architectures are important, the software ecosystem often dictates project success. Ultralytics provides a [streamlined user experience](https://docs.ultralytics.com/usage/python/) and robust deployment tools that significantly reduce time-to-market compared to complex, research-oriented codebases. + While theoretical architectures are important, the software ecosystem often dictates project success. Ultralytics provides a [streamlined user experience](https://docs.ultralytics.com/usage/python) and robust deployment tools that significantly reduce time-to-market compared to complex, research-oriented codebases. ## Performance and Metrics Comparison @@ -73,19 +73,19 @@ When analyzing model performance, balancing precision with [inference latency](h ### Critical Analysis of Metrics 1. **Accuracy Thresholds:** YOLOv9e achieves the highest overall accuracy at an impressive 55.6% [mAP (mean Average Precision)](https://www.ultralytics.com/glossary/mean-average-precision-map), outperforming the heaviest EfficientDet-d7 model (53.7%) while maintaining faster TensorRT speeds. -2. **Real-Time Speed:** YOLOv9t requires only 2.3ms on a T4 GPU using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), emphasizing the efficiency of the GELAN architecture for high-speed video streams. EfficientDet-d0 operates rapidly but sacrifices significant mAP to reach those speeds. +2. **Real-Time Speed:** YOLOv9t requires only 2.3ms on a T4 GPU using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), emphasizing the efficiency of the GELAN architecture for high-speed video streams. EfficientDet-d0 operates rapidly but sacrifices significant mAP to reach those speeds. 3. **Computational Complexity:** EfficientDet scales heavily in parameter count and FLOPs as the compound factor increases. The d7 variant reaches 128ms latency, making it over 10x slower than comparable modern YOLO models, heavily restricting its use in [real-time inference](https://www.ultralytics.com/glossary/real-time-inference) environments. ## Training Efficiency and Ecosystem -Choosing a model involves evaluating the developer ecosystem. The [Ultralytics ecosystem](https://docs.ultralytics.com/integrations/) provides an unparalleled advantage in training efficiency, deployment flexibility, and general versatility. +Choosing a model involves evaluating the developer ecosystem. The [Ultralytics ecosystem](https://docs.ultralytics.com/integrations) provides an unparalleled advantage in training efficiency, deployment flexibility, and general versatility. ### The Ultralytics Advantage Models supported within the Ultralytics framework, including YOLOv9 through community integrations and official Ultralytics models like YOLOv8 and YOLO11, benefit from dramatically lower memory requirements during training compared to transformer-based or older TensorFlow architectures like EfficientDet. The robust PyTorch backend ensures fast convergence and stability. -- **Versatility:** Unlike EfficientDet, which strictly focuses on bounding box detection, the Ultralytics API natively supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). -- **Ease of Use:** EfficientDet relies on older TensorFlow libraries and complex AutoML configurations, which can be brittle to set up. In contrast, Ultralytics offers a highly refined API for seamless [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/) and dataset management. +- **Versatility:** Unlike EfficientDet, which strictly focuses on bounding box detection, the Ultralytics API natively supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Image Classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). +- **Ease of Use:** EfficientDet relies on older TensorFlow libraries and complex AutoML configurations, which can be brittle to set up. In contrast, Ultralytics offers a highly refined API for seamless [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning) and dataset management. ### Implementation Example @@ -120,7 +120,7 @@ While YOLOv9 and EfficientDet are powerful, developers looking for the ultimate Released in January 2026, **[Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26)** represents the current state-of-the-art. It improves upon previous generations (including [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8)) with several critical breakthroughs: -- **End-to-End NMS-Free Design:** YOLO26 eliminates Non-Maximum Suppression entirely, a concept pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), resulting in significantly faster and simpler [model deployment](https://docs.ultralytics.com/guides/model-deployment-options/). +- **End-to-End NMS-Free Design:** YOLO26 eliminates Non-Maximum Suppression entirely, a concept pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), resulting in significantly faster and simpler [model deployment](https://docs.ultralytics.com/guides/model-deployment-options). - **DFL Removal:** Distribution Focal Loss removed for simplified export and better edge/low-power device compatibility. - **Up to 43% Faster CPU Inference:** Perfectly optimized for [IoT devices](https://www.ultralytics.com/blog/industrial-iot-iiot-internet-of-things-explained) and environments lacking dedicated GPUs. - **MuSGD Optimizer:** A revolutionary hybrid of SGD and Muon (inspired by LLM training innovations), ensuring faster convergence and incredibly stable training runs. diff --git a/docs/en/compare/yolov9-vs-pp-yoloe.md b/docs/en/compare/yolov9-vs-pp-yoloe.md index a7da0b1d01d..ca5d4bdd2f8 100644 --- a/docs/en/compare/yolov9-vs-pp-yoloe.md +++ b/docs/en/compare/yolov9-vs-pp-yoloe.md @@ -6,7 +6,7 @@ keywords: YOLOv9,PP-YOLOE+,object detection,model comparison,computer vision,AI, # YOLOv9 vs. PP-YOLOE+: A Technical Deep Dive into Modern Object Detection -The landscape of real-time object detection continues to advance rapidly, offering computer vision engineers a wide array of choices for deploying highly accurate models on edge and cloud infrastructure. Two prominent models in this space are **[YOLOv9](https://docs.ultralytics.com/models/yolov9/)** and **PP-YOLOE+**. While both push the boundaries of accuracy and speed, they emerge from different research lineages and software ecosystems. +The landscape of real-time object detection continues to advance rapidly, offering computer vision engineers a wide array of choices for deploying highly accurate models on edge and cloud infrastructure. Two prominent models in this space are **[YOLOv9](https://docs.ultralytics.com/models/yolov9)** and **PP-YOLOE+**. While both push the boundaries of accuracy and speed, they emerge from different research lineages and software ecosystems. This comprehensive technical comparison explores their architectures, training methodologies, performance metrics, and ideal real-world applications. We will also explore how the broader [Ultralytics ecosystem](https://www.ultralytics.com) provides significant advantages for developers prioritizing ease of use, memory efficiency, and versatile deployment. @@ -28,9 +28,9 @@ Introduced in early 2024, YOLOv9 tackles the data loss that occurs as informatio - **Date:** February 21, 2024 - **Arxiv:** [2402.13616](https://arxiv.org/abs/2402.13616) - **GitHub:** [WongKinYiu/yolov9](https://github.com/WongKinYiu/yolov9) -- **Docs:** [Ultralytics YOLOv9 Documentation](https://docs.ultralytics.com/models/yolov9/) +- **Docs:** [Ultralytics YOLOv9 Documentation](https://docs.ultralytics.com/models/yolov9) -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ### PP-YOLOE+: Advancing the Paddle Ecosystem @@ -76,7 +76,7 @@ When evaluating models for production, the trade-off between mAP (mean Average P ### Analysis - **Parameter Efficiency:** YOLOv9 achieves remarkably higher efficiency. For instance, YOLOv9c reaches an mAP of 53.0% using only 25.3M parameters, while PP-YOLOE+l requires over double the parameters (52.2M) to achieve a slightly lower mAP of 52.9%. This drastically lowers the memory requirements for YOLOv9. -- **Inference Speed:** YOLOv9 models demonstrate excellent optimization for hardware accelerators like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), yielding competitive inference speeds on NVIDIA T4 GPUs that are crucial for [real-time inference](https://www.ultralytics.com/blog/real-time-inferences-in-vision-ai-solutions-are-making-an-impact). +- **Inference Speed:** YOLOv9 models demonstrate excellent optimization for hardware accelerators like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), yielding competitive inference speeds on NVIDIA T4 GPUs that are crucial for [real-time inference](https://www.ultralytics.com/blog/real-time-inferences-in-vision-ai-solutions-are-making-an-impact). ## Training Methodologies and Ecosystem @@ -88,7 +88,7 @@ PP-YOLOE+ is tightly coupled with the [PaddleDetection](https://github.com/Paddl ### The Ultralytics Advantage: Streamlined Workflows -In contrast, YOLOv9 operates within the highly polished **Ultralytics ecosystem**. Designed for developers and researchers, Ultralytics prioritizes an exceptional ease of use. The [Python API](https://docs.ultralytics.com/usage/python/) completely abstracts away complex boilerplate code. +In contrast, YOLOv9 operates within the highly polished **Ultralytics ecosystem**. Designed for developers and researchers, Ultralytics prioritizes an exceptional ease of use. The [Python API](https://docs.ultralytics.com/usage/python) completely abstracts away complex boilerplate code. ```python from ultralytics import YOLO @@ -106,7 +106,7 @@ results = model("https://ultralytics.com/images/bus.jpg") model.export(format="onnx") ``` -This workflow highlights the superior **Training Efficiency** of Ultralytics models. Native support for data augmentation, distributed training, and automatic logging to platforms like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) or [MLflow](https://docs.ultralytics.com/integrations/mlflow/) comes standard. +This workflow highlights the superior **Training Efficiency** of Ultralytics models. Native support for data augmentation, distributed training, and automatic logging to platforms like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) or [MLflow](https://docs.ultralytics.com/integrations/mlflow) comes standard. !!! tip "Explore the Latest in Vision AI" @@ -118,7 +118,7 @@ Modern computer vision projects rarely stop at simple bounding boxes. PP-YOLOE+ is primarily engineered for standard object detection. Adapting its architecture for other tasks involves extensive custom engineering. -Conversely, the Ultralytics framework is a multi-task powerhouse. By utilizing a unified API, developers can effortlessly switch from standard object detection to complex [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), highly accurate [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection for aerial imagery, and Image [Classification](https://docs.ultralytics.com/tasks/classify/). This unparalleled versatility is why enterprise teams consistently choose Ultralytics models like YOLOv9, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), and YOLO26. +Conversely, the Ultralytics framework is a multi-task powerhouse. By utilizing a unified API, developers can effortlessly switch from standard object detection to complex [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), highly accurate [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection for aerial imagery, and Image [Classification](https://docs.ultralytics.com/tasks/classify). This unparalleled versatility is why enterprise teams consistently choose Ultralytics models like YOLOv9, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), and YOLO26. ## Ideal Use Cases and Applications @@ -126,4 +126,4 @@ Conversely, the Ultralytics framework is a multi-task powerhouse. By utilizing a - **Retail Inventory Systems:** For detecting dense configurations of small items on shelves, YOLOv9's PGI effectively maintains fine-grained spatial details, outperforming PP-YOLOE+ on small-object detection tasks. - **Legacy Deployments:** **PP-YOLOE+** remains a viable option strictly for teams explicitly mandated to use the Baidu/PaddlePaddle software stack in existing legacy infrastructure. -For researchers exploring Transformer-based architectures, Ultralytics also natively supports **[RT-DETR](https://docs.ultralytics.com/models/rtdetr/)** within the exact same easy-to-use API, ensuring you always have access to the optimal model for your specific deployment requirements. +For researchers exploring Transformer-based architectures, Ultralytics also natively supports **[RT-DETR](https://docs.ultralytics.com/models/rtdetr)** within the exact same easy-to-use API, ensuring you always have access to the optimal model for your specific deployment requirements. diff --git a/docs/en/compare/yolov9-vs-rtdetr.md b/docs/en/compare/yolov9-vs-rtdetr.md index ac6754ec5b3..c4159ed9aff 100644 --- a/docs/en/compare/yolov9-vs-rtdetr.md +++ b/docs/en/compare/yolov9-vs-rtdetr.md @@ -6,9 +6,9 @@ keywords: YOLOv9, RTDETRv2, object detection, model comparison, AI models, compu # YOLOv9 vs. RTDETRv2: A Technical Deep Dive into Modern Object Detection -The landscape of real-time [object detection](https://docs.ultralytics.com/tasks/detect/) has experienced a paradigm shift in recent years. Two distinct architectural philosophies have emerged to dominate the field: highly optimized Convolutional Neural Networks (CNNs) and real-time Detection Transformers (DETRs). Representing the pinnacle of these two approaches are **YOLOv9** and **RTDETRv2**. +The landscape of real-time [object detection](https://docs.ultralytics.com/tasks/detect) has experienced a paradigm shift in recent years. Two distinct architectural philosophies have emerged to dominate the field: highly optimized Convolutional Neural Networks (CNNs) and real-time Detection Transformers (DETRs). Representing the pinnacle of these two approaches are **YOLOv9** and **RTDETRv2**. -This comprehensive guide compares these two powerful models, analyzing their architectural innovations, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/), and ideal deployment scenarios to help you choose the right model for your [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) pipeline. +This comprehensive guide compares these two powerful models, analyzing their architectural innovations, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics), and ideal deployment scenarios to help you choose the right model for your [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) pipeline. @@ -35,7 +35,7 @@ Understanding the origins and design intent of these models provides crucial con **Arxiv:** [https://arxiv.org/abs/2402.13616](https://arxiv.org/abs/2402.13616) **GitHub:** [WongKinYiu/yolov9](https://github.com/WongKinYiu/yolov9) -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ### RTDETRv2 @@ -45,20 +45,20 @@ Understanding the origins and design intent of these models provides crucial con **Arxiv:** [https://arxiv.org/abs/2407.17140](https://arxiv.org/abs/2407.17140) **GitHub:** [lyuwenyu/RT-DETR](https://github.com/lyuwenyu/RT-DETR/tree/main/rtdetrv2_pytorch) -[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr){ .md-button } ## Architectural Innovations ### YOLOv9: Solving the Information Bottleneck -[Ultralytics YOLOv9](https://docs.ultralytics.com/models/yolov9/) introduces two major innovations designed to address information loss as data passes through deep neural networks: +[Ultralytics YOLOv9](https://docs.ultralytics.com/models/yolov9) introduces two major innovations designed to address information loss as data passes through deep neural networks: 1. **Programmable Gradient Information (PGI):** This auxiliary supervision framework ensures that reliable gradients are generated to update network weights, preserving crucial feature information even in very deep network layers. 2. **Generalized Efficient Layer Aggregation Network (GELAN):** A novel architecture that combines the strengths of CSPNet and ELAN. GELAN optimizes parameter efficiency, allowing YOLOv9 to achieve higher accuracy with fewer FLOPs compared to traditional CNNs. ### RTDETRv2: Enhancing Real-Time Transformers -Building upon the success of the original RT-DETR, [RTDETRv2](https://docs.ultralytics.com/models/rtdetr/) utilizes a Transformer-based architecture that inherently avoids the need for Non-Maximum Suppression (NMS). Its improvements include: +Building upon the success of the original RT-DETR, [RTDETRv2](https://docs.ultralytics.com/models/rtdetr) utilizes a Transformer-based architecture that inherently avoids the need for Non-Maximum Suppression (NMS). Its improvements include: 1. **Bag-of-Freebies Strategy:** The v2 iteration incorporates advanced training techniques and data augmentations that significantly boost accuracy without adding any overhead to inference latency. 2. **Efficient Hybrid Encoder:** By processing multi-scale features through a decoupled intra-scale and cross-scale attention mechanism, RTDETRv2 efficiently manages the traditionally high computational cost of Vision Transformers. @@ -92,7 +92,7 @@ Conversely, **RTDETRv2** provides strong competition in the medium-to-large mode ## The Ultralytics Advantage: Ecosystem and Versatility -While pure architectural metrics are important, the software ecosystem often dictates the success of an AI project. Accessing these advanced models through the [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) offers unparalleled advantages. +While pure architectural metrics are important, the software ecosystem often dictates the success of an AI project. Accessing these advanced models through the [Ultralytics Python API](https://docs.ultralytics.com/usage/python) offers unparalleled advantages. ### Streamlined Training and Deployment @@ -115,13 +115,13 @@ model_yolo.export(format="openvino") ### Unmatched Task Versatility -A major limitation of specialized models like RTDETRv2 is their narrow focus on bounding box detection. In contrast, the broader Ultralytics ecosystem, encompassing models like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), supports a wide array of [computer vision tasks](https://docs.ultralytics.com/tasks/). This includes pixel-perfect [instance segmentation](https://docs.ultralytics.com/tasks/segment/), skeletal [pose estimation](https://docs.ultralytics.com/tasks/pose/), whole-image [classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection for aerial imagery. +A major limitation of specialized models like RTDETRv2 is their narrow focus on bounding box detection. In contrast, the broader Ultralytics ecosystem, encompassing models like [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8), supports a wide array of [computer vision tasks](https://docs.ultralytics.com/tasks). This includes pixel-perfect [instance segmentation](https://docs.ultralytics.com/tasks/segment), skeletal [pose estimation](https://docs.ultralytics.com/tasks/pose), whole-image [classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection for aerial imagery. ## Real-World Applications ### High-Speed Edge Analytics -For retail environments or manufacturing lines requiring real-time product recognition on edge devices, **YOLOv9** is the superior choice. Its [GELAN architecture](https://docs.ultralytics.com/models/yolov9/) ensures high throughput on constrained hardware like the NVIDIA Jetson series, enabling automated quality control without significant lag. +For retail environments or manufacturing lines requiring real-time product recognition on edge devices, **YOLOv9** is the superior choice. Its [GELAN architecture](https://docs.ultralytics.com/models/yolov9) ensures high throughput on constrained hardware like the NVIDIA Jetson series, enabling automated quality control without significant lag. ### Complex Scene Analysis @@ -149,16 +149,16 @@ RT-DETR is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Future: Enter YOLO26 While YOLOv9 and RTDETRv2 represent massive achievements, the [computer vision](https://www.ultralytics.com/glossary/computer-vision-cv) field moves rapidly. For developers looking to start new projects, **[YOLO26](https://platform.ultralytics.com/ultralytics/yolo26)** is the recommended state-of-the-art solution. -Released in 2026, YOLO26 incorporates the best features of both CNNs and DETRs. It features an **End-to-End NMS-Free Design**, completely eliminating post-processing latency—a technique first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/). Furthermore, YOLO26 removes Distribution Focal Loss (DFL) for better edge compatibility and introduces the revolutionary **MuSGD Optimizer**. Inspired by Large Language Model training (specifically Moonshot AI's Kimi K2), this hybrid optimizer ensures unprecedented training stability and faster convergence. +Released in 2026, YOLO26 incorporates the best features of both CNNs and DETRs. It features an **End-to-End NMS-Free Design**, completely eliminating post-processing latency—a technique first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10). Furthermore, YOLO26 removes Distribution Focal Loss (DFL) for better edge compatibility and introduces the revolutionary **MuSGD Optimizer**. Inspired by Large Language Model training (specifically Moonshot AI's Kimi K2), this hybrid optimizer ensures unprecedented training stability and faster convergence. Coupled with improved loss functions like ProgLoss and STAL for exceptional small-object recognition, YOLO26 delivers up to **43% faster CPU inference**, solidifying its position as the ultimate model for modern AI deployments. diff --git a/docs/en/compare/yolov9-vs-yolo11.md b/docs/en/compare/yolov9-vs-yolo11.md index 851eef1d507..9994705fc62 100644 --- a/docs/en/compare/yolov9-vs-yolo11.md +++ b/docs/en/compare/yolov9-vs-yolo11.md @@ -6,7 +6,7 @@ keywords: YOLO11, YOLOv9, object detection, model comparison, benchmarks, Ultral # YOLOv9 vs. YOLO11: A Technical Deep Dive into Modern Object Detection -The rapid evolution of computer vision has continuously pushed the boundaries of what is possible in real-time [object detection](https://docs.ultralytics.com/tasks/detect/). When comparing leading architectures, **YOLOv9** and **[Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11)** stand out as monumental leaps forward, each serving distinct technical needs. YOLOv9 introduced novel ways to preserve gradient flow during deep network training, while YOLO11 revolutionized the general-purpose vision ecosystem with unmatched efficiency, versatility, and ease of use. +The rapid evolution of computer vision has continuously pushed the boundaries of what is possible in real-time [object detection](https://docs.ultralytics.com/tasks/detect). When comparing leading architectures, **YOLOv9** and **[Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11)** stand out as monumental leaps forward, each serving distinct technical needs. YOLOv9 introduced novel ways to preserve gradient flow during deep network training, while YOLO11 revolutionized the general-purpose vision ecosystem with unmatched efficiency, versatility, and ease of use. This comprehensive technical comparison analyzes their architectures, performance metrics, memory requirements, and ideal deployment scenarios to help you select the optimal model for your next AI project. @@ -33,7 +33,7 @@ YOLOv9 brought a strong academic focus on deep learning information bottlenecks, - **Arxiv:** [https://arxiv.org/abs/2402.13616](https://arxiv.org/abs/2402.13616) - **GitHub:** [https://github.com/WongKinYiu/yolov9](https://github.com/WongKinYiu/yolov9) -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ### Ultralytics YOLO11 @@ -54,13 +54,13 @@ YOLOv9 introduces the concept of Programmable Gradient Information (PGI) alongsi ### Streamlined Efficiency in YOLO11 -YOLO11 builds on years of foundational research to deliver a highly optimized architecture. It improves upon previous iterations by reducing computational overhead while maximizing feature extraction. Unlike traditional NMS pipelines that bottleneck CPU performance, YOLO11 uses refined detection heads that achieve an incredible balance between latency and precision. Furthermore, YOLO11 boasts inherently lower memory usage during both [model training](https://docs.ultralytics.com/modes/train/) and inference compared to heavy [Transformer](https://huggingface.co/docs/transformers/index) models, which are often slower to train and require massive amounts of CUDA memory. +YOLO11 builds on years of foundational research to deliver a highly optimized architecture. It improves upon previous iterations by reducing computational overhead while maximizing feature extraction. Unlike traditional NMS pipelines that bottleneck CPU performance, YOLO11 uses refined detection heads that achieve an incredible balance between latency and precision. Furthermore, YOLO11 boasts inherently lower memory usage during both [model training](https://docs.ultralytics.com/modes/train) and inference compared to heavy [Transformer](https://huggingface.co/docs/transformers/index) models, which are often slower to train and require massive amounts of CUDA memory. ## Performance Metrics Comparison When comparing these models on the standard [COCO dataset](https://cocodataset.org/), both showcase incredible capabilities, but trade-offs emerge between raw parameter count and operational speed. -Below is a detailed breakdown of [YOLO Performance Metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/). +Below is a detailed breakdown of [YOLO Performance Metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -92,14 +92,14 @@ The original YOLOv9 repository is highly specialized, offering cutting-edge rese YOLOv9 focuses predominantly on bounding box detection. In contrast, YOLO11 is a unified multi-task powerhouse natively supporting: -- [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/) -- [Pose Estimation](https://docs.ultralytics.com/tasks/pose/) -- [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) -- [Image Classification](https://docs.ultralytics.com/tasks/classify/) +- [Instance Segmentation](https://docs.ultralytics.com/tasks/segment) +- [Pose Estimation](https://docs.ultralytics.com/tasks/pose) +- [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb) +- [Image Classification](https://docs.ultralytics.com/tasks/classify) ### Seamless Deployment -Using the Ultralytics ecosystem allows developers to seamlessly [export models](https://docs.ultralytics.com/modes/export/) to an array of formats with a single line of [Python](https://www.python.org/) code. Whether targeting [ONNX](https://onnx.ai/), [OpenVINO](https://docs.openvino.ai/), [TFLite](https://ai.google.dev/edge/litert), or [CoreML](https://developer.apple.com/machine-learning/core-ml/), the transition from training to production is effortless. +Using the Ultralytics ecosystem allows developers to seamlessly [export models](https://docs.ultralytics.com/modes/export) to an array of formats with a single line of [Python](https://www.python.org/) code. Whether targeting [ONNX](https://onnx.ai/), [OpenVINO](https://docs.openvino.ai/), [TFLite](https://ai.google.dev/edge/litert), or [CoreML](https://developer.apple.com/machine-learning/core-ml/), the transition from training to production is effortless. ```python from ultralytics import YOLO @@ -126,10 +126,10 @@ For developers, engineers, and production teams, **YOLO11 is highly recommended* - **Smart Retail Analytics:** Tracking products and customers seamlessly using standard [Intel standard processors](https://www.intel.com/). - **Autonomous Drones:** Where low-FLOP architectures preserve battery life while still delivering robust small-object detection. -- **Dynamic Projects:** Workflows that might start as detection but evolve to require [pose estimation](https://docs.ultralytics.com/tasks/pose/) or segmentation later on. +- **Dynamic Projects:** Workflows that might start as detection but evolve to require [pose estimation](https://docs.ultralytics.com/tasks/pose) or segmentation later on. ## Looking Ahead: The Next Evolution While YOLO11 represents the state-of-the-art for its generation, the computer vision landscape continues to advance. Users exploring the boundaries of AI should also look toward **[YOLO26](https://platform.ultralytics.com/ultralytics/yolo26)**. -Pioneering an end-to-end NMS-free design first explored in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 introduces the MuSGD optimizer (a hybrid of SGD and Muon) for unprecedented training stability. With the removal of Distribution Focal Loss (DFL) to simplify export, and advanced loss mechanisms like ProgLoss and STAL, YOLO26 achieves up to 43% faster CPU inference. For modern projects, it offers the ultimate combination of academic innovation and production-ready reliability. Furthermore, teams upgrading from legacy systems like [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) will find the transition to YOLO26 or YOLO11 entirely frictionless thanks to the unified Ultralytics API. +Pioneering an end-to-end NMS-free design first explored in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 introduces the MuSGD optimizer (a hybrid of SGD and Muon) for unprecedented training stability. With the removal of Distribution Focal Loss (DFL) to simplify export, and advanced loss mechanisms like ProgLoss and STAL, YOLO26 achieves up to 43% faster CPU inference. For modern projects, it offers the ultimate combination of academic innovation and production-ready reliability. Furthermore, teams upgrading from legacy systems like [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) will find the transition to YOLO26 or YOLO11 entirely frictionless thanks to the unified Ultralytics API. diff --git a/docs/en/compare/yolov9-vs-yolo26.md b/docs/en/compare/yolov9-vs-yolo26.md index 7f78d6c1ac5..d5877cfb4d0 100644 --- a/docs/en/compare/yolov9-vs-yolo26.md +++ b/docs/en/compare/yolov9-vs-yolo26.md @@ -21,11 +21,11 @@ Understanding the origins of these deep learning models provides valuable contex Authored by Chien-Yao Wang and Hong-Yuan Mark Liao from the Institute of Information Science at [Academia Sinica](https://www.iis.sinica.edu.tw/en/index.html) in Taiwan, YOLOv9 was released on February 21, 2024. The model focuses heavily on theoretical deep learning concepts, specifically addressing the information bottleneck problem in deep convolutional neural networks (CNNs). -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ### Ultralytics YOLO26 -Authored by Glenn Jocher and Jing Qiu at [Ultralytics](https://www.ultralytics.com/), YOLO26 was released on January 14, 2026. Building on the massive success of predecessors like [YOLO11](https://docs.ultralytics.com/models/yolo11/) and [YOLOv8](https://docs.ultralytics.com/models/yolov8/), YOLO26 was engineered from the ground up to prioritize production readiness, edge deployment, and native end-to-end efficiency. +Authored by Glenn Jocher and Jing Qiu at [Ultralytics](https://www.ultralytics.com/), YOLO26 was released on January 14, 2026. Building on the massive success of predecessors like [YOLO11](https://docs.ultralytics.com/models/yolo11) and [YOLOv8](https://docs.ultralytics.com/models/yolov8), YOLO26 was engineered from the ground up to prioritize production readiness, edge deployment, and native end-to-end efficiency. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -45,9 +45,9 @@ However, YOLOv9 relies heavily on traditional Non-Maximum Suppression (NMS) for ### The Edge-First Architecture of YOLO26 -YOLO26 takes a radically different approach by optimizing the entire pipeline from training to real-time deployment. It builds upon the **End-to-End NMS-Free Design** first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), entirely eliminating the need for NMS post-processing. This results in incredibly low latency, making it heavily optimized for edge devices like the [Raspberry Pi](https://www.raspberrypi.org/) or [NVIDIA Jetson](https://developer.nvidia.com/embedded/jetson-nano). +YOLO26 takes a radically different approach by optimizing the entire pipeline from training to real-time deployment. It builds upon the **End-to-End NMS-Free Design** first pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), entirely eliminating the need for NMS post-processing. This results in incredibly low latency, making it heavily optimized for edge devices like the [Raspberry Pi](https://www.raspberrypi.org/) or [NVIDIA Jetson](https://developer.nvidia.com/embedded/jetson-nano). -Furthermore, YOLO26 completely removes Distribution Focal Loss (DFL). This structural change simplifies model [exporting to ONNX](https://docs.ultralytics.com/integrations/onnx/) and provides significantly better compatibility with low-power microcontrollers. +Furthermore, YOLO26 completely removes Distribution Focal Loss (DFL). This structural change simplifies model [exporting to ONNX](https://docs.ultralytics.com/integrations/onnx) and provides significantly better compatibility with low-power microcontrollers. For the training phase, YOLO26 integrates the novel **MuSGD Optimizer**, a hybrid of [Stochastic Gradient Descent](https://en.wikipedia.org/wiki/Stochastic_gradient_descent) and Muon (inspired by the LLM training methodologies of Moonshot AI's Kimi K2). This bridges the gap between Large Language Model (LLM) training innovations and computer vision, offering drastically more stable training and faster convergence times. @@ -73,7 +73,7 @@ When benchmarking on the widely used [COCO dataset](https://cocodataset.org/), b - **Speed and Efficiency:** Because YOLO26 utilizes an NMS-free architecture and simplified loss functions, it boasts **up to 43% faster CPU inference** compared to legacy architectures. The YOLO26n model runs at a blistering 1.7ms on an NVIDIA T4 GPU using [TensorRT](https://developer.nvidia.com/tensorrt), making it the ultimate choice for real-time video streams. - **Accuracy:** The YOLO26x model achieves an unparalleled **57.5 mAP**, outperforming the largest YOLOv9e model while maintaining lower latency. -- **Memory Requirements:** Ultralytics models are known for their efficiency. YOLO26 requires significantly less CUDA memory during [model training](https://docs.ultralytics.com/modes/train/) and inference compared to complex [transformer-based vision models](https://en.wikipedia.org/wiki/Vision_transformer), allowing developers to utilize larger batch sizes on consumer-grade hardware. +- **Memory Requirements:** Ultralytics models are known for their efficiency. YOLO26 requires significantly less CUDA memory during [model training](https://docs.ultralytics.com/modes/train) and inference compared to complex [transformer-based vision models](https://en.wikipedia.org/wiki/Vision_transformer), allowing developers to utilize larger batch sizes on consumer-grade hardware. ## Ecosystem, Ease of Use, and Versatility @@ -98,12 +98,12 @@ model.export(format="onnx") ### Unmatched Task Versatility -Unlike YOLOv9, which is primarily tailored for standard object detection, YOLO26 natively supports a vast array of [computer vision tasks](https://docs.ultralytics.com/tasks/) out of the box. The architecture includes specific enhancements for diverse applications: +Unlike YOLOv9, which is primarily tailored for standard object detection, YOLO26 natively supports a vast array of [computer vision tasks](https://docs.ultralytics.com/tasks) out of the box. The architecture includes specific enhancements for diverse applications: -- **[Instance Segmentation](https://docs.ultralytics.com/tasks/segment/):** Features a specialized semantic segmentation loss and multi-scale proto for flawless pixel-level masks. -- **[Pose Estimation](https://docs.ultralytics.com/tasks/pose/):** Integrates Residual Log-Likelihood Estimation (RLE) to track skeletal keypoints with extreme precision. -- **[Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/):** Includes a specialized angle loss function designed specifically to solve boundary issues in rotated object detection for aerial imagery. -- **[Image Classification](https://docs.ultralytics.com/tasks/classify/):** Robust categorization for entire images based on [ImageNet](https://www.image-net.org/) standards. +- **[Instance Segmentation](https://docs.ultralytics.com/tasks/segment):** Features a specialized semantic segmentation loss and multi-scale proto for flawless pixel-level masks. +- **[Pose Estimation](https://docs.ultralytics.com/tasks/pose):** Integrates Residual Log-Likelihood Estimation (RLE) to track skeletal keypoints with extreme precision. +- **[Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb):** Includes a specialized angle loss function designed specifically to solve boundary issues in rotated object detection for aerial imagery. +- **[Image Classification](https://docs.ultralytics.com/tasks/classify):** Robust categorization for entire images based on [ImageNet](https://www.image-net.org/) standards. !!! info "Integrated Ecosystem" @@ -123,7 +123,7 @@ For robotics, autonomous drones, and smart home IoT devices, **YOLO26 is the und ### High-Speed Manufacturing Pipelines -In industrial settings like automated [defect detection](https://docs.ultralytics.com/guides/object-cropping/) on high-speed conveyor belts, the blazing-fast TensorRT speeds of YOLO26 models ensure that no frames are dropped, maximizing the throughput of quality assurance systems. +In industrial settings like automated [defect detection](https://docs.ultralytics.com/guides/object-cropping) on high-speed conveyor belts, the blazing-fast TensorRT speeds of YOLO26 models ensure that no frames are dropped, maximizing the throughput of quality assurance systems. ## Use Cases and Recommendations @@ -143,7 +143,7 @@ YOLO26 is recommended for: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Conclusion diff --git a/docs/en/compare/yolov9-vs-yolov10.md b/docs/en/compare/yolov9-vs-yolov10.md index 3f2d0d8d179..0cfaa56f358 100644 --- a/docs/en/compare/yolov9-vs-yolov10.md +++ b/docs/en/compare/yolov9-vs-yolov10.md @@ -8,7 +8,7 @@ keywords: YOLOv9, YOLOv10, object detection, Ultralytics, computer vision, model The landscape of real-time computer vision has seen immense advancements, driven largely by researchers continuously pushing the performance-efficiency boundary. When analyzing the evolution of state-of-the-art vision models, **YOLOv9** and **YOLOv10** represent two critical milestones. Released in early 2024, both models introduced paradigm-shifting architectural designs to address long-standing challenges in deep neural networks, from information bottlenecks to post-processing latency. -This comprehensive technical comparison explores their architectures, performance metrics, and ideal deployment scenarios, helping you navigate the complexities of modern [object detection](https://docs.ultralytics.com/tasks/detect/) ecosystems. +This comprehensive technical comparison explores their architectures, performance metrics, and ideal deployment scenarios, helping you navigate the complexities of modern [object detection](https://docs.ultralytics.com/tasks/detect) ecosystems. @@ -30,7 +30,7 @@ Introduced on February 21, 2024, YOLOv9 tackles the theoretical issue of informa YOLOv9 introduces the **Generalized Efficient Layer Aggregation Network (GELAN)**, which maximizes parameter utilization by combining the strengths of CSPNet and ELAN. Furthermore, it employs **Programmable Gradient Information (PGI)**, an auxiliary supervision mechanism ensuring deep layers retain critical spatial information. This makes YOLOv9 exceptionally strong for tasks demanding high feature fidelity, such as [medical image analysis](https://www.ultralytics.com/glossary/medical-image-analysis) or distant surveillance. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ### YOLOv10: Real-Time End-to-End Efficiency @@ -43,7 +43,7 @@ Released shortly after on May 23, 2024, YOLOv10 reimagines the deployment pipeli YOLOv10 utilizes **consistent dual assignments** during training, allowing for a natively **NMS-free design**. This removes post-processing overhead during inference, drastically reducing latency. Combined with a holistic efficiency-accuracy driven model design, YOLOv10 achieves an outstanding balance, lowering computational overhead (FLOPs) while maintaining competitive precision, making it highly attractive for [edge computing](https://www.ultralytics.com/glossary/edge-computing) applications. -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## Performance and Metrics Comparison @@ -82,7 +82,7 @@ While architectural differences are critical, the surrounding software ecosystem Unlike complex transformer-based architectures that suffer from massive memory bloat, Ultralytics YOLO models are engineered for optimal [GPU memory](https://www.ultralytics.com/glossary/gpu-graphics-processing-unit) usage. This allows researchers to utilize larger [batch sizes](https://www.ultralytics.com/glossary/batch-size) on consumer-grade hardware, making state-of-the-art AI accessible. -The unified Python API abstracts away the complexities of [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation/) and [hyperparameter tuning](https://www.ultralytics.com/glossary/hyperparameter-tuning). You can seamlessly switch between architectures simply by altering the weight file string. +The unified Python API abstracts away the complexities of [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation) and [hyperparameter tuning](https://www.ultralytics.com/glossary/hyperparameter-tuning). You can seamlessly switch between architectures simply by altering the weight file string. ```python from ultralytics import YOLO @@ -100,7 +100,7 @@ metrics = model.val() model.export(format="onnx") ``` -Whether you need to log metrics to [MLflow](https://docs.ultralytics.com/integrations/mlflow/) or export to [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) for high-speed hardware deployment, the Ultralytics platform handles it natively. +Whether you need to log metrics to [MLflow](https://docs.ultralytics.com/integrations/mlflow) or export to [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) for high-speed hardware deployment, the Ultralytics platform handles it natively. ## Ideal Use Cases @@ -120,7 +120,7 @@ YOLO26 represents the ultimate synthesis of previous generations, combining the - **End-to-End NMS-Free Design:** Building on the foundations laid by YOLOv10, YOLO26 natively eliminates NMS post-processing for simpler deployment. - **MuSGD Optimizer:** A hybrid of SGD and Muon, bringing advanced LLM training innovations to computer vision for incredibly stable and fast convergence. - **Up to 43% Faster CPU Inference:** Specifically optimized for edge computing and devices without dedicated GPUs. -- **DFL Removal:** Distribution Focal Loss was removed to simplify [model export](https://docs.ultralytics.com/modes/export/) and boost low-power device compatibility. +- **DFL Removal:** Distribution Focal Loss was removed to simplify [model export](https://docs.ultralytics.com/modes/export) and boost low-power device compatibility. - **ProgLoss + STAL:** These improved loss functions bring notable improvements in small-object recognition, matching or exceeding YOLOv9's capabilities. -For researchers evaluating legacy architectures, [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) and [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) are also well-documented alternatives within the Ultralytics ecosystem. However, for maximum versatility across all vision tasks, transitioning to YOLO26 on the [Ultralytics Platform](https://platform.ultralytics.com) ensures you are leveraging the pinnacle of open-source vision AI. +For researchers evaluating legacy architectures, [RT-DETR](https://docs.ultralytics.com/models/rtdetr) and [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) are also well-documented alternatives within the Ultralytics ecosystem. However, for maximum versatility across all vision tasks, transitioning to YOLO26 on the [Ultralytics Platform](https://platform.ultralytics.com) ensures you are leveraging the pinnacle of open-source vision AI. diff --git a/docs/en/compare/yolov9-vs-yolov5.md b/docs/en/compare/yolov9-vs-yolov5.md index 08514a8bd41..5463a0ab000 100644 --- a/docs/en/compare/yolov9-vs-yolov5.md +++ b/docs/en/compare/yolov9-vs-yolov5.md @@ -23,13 +23,13 @@ Developed by researchers Chien-Yao Wang and Hong-Yuan Mark Liao at the Institute By utilizing PGI, YOLOv9 ensures that vital information is retained throughout the feed-forward process, leading to highly accurate gradient updates. Meanwhile, the GELAN architecture maximizes parameter efficiency, allowing the model to achieve state-of-the-art accuracy with surprisingly low computational overhead. You can explore the technical details in the official [YOLOv9 Arxiv paper](https://arxiv.org/abs/2402.13616) or view the [YOLOv9 GitHub repository](https://github.com/WongKinYiu/yolov9). -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ### Ultralytics YOLOv5: The Production Standard Authored by Glenn Jocher and released by Ultralytics on June 26, 2020, YOLOv5 revolutionized the accessibility of computer vision. As one of the first object detection models built natively on the [PyTorch](https://pytorch.org/) framework, it bypassed the complexities of the older Darknet C-framework. YOLOv5 leverages a highly optimized CSPNet backbone and a PANet neck, prioritizing a seamless balance between speed and accuracy. -Its crowning achievement, however, is its integration into the broader Ultralytics ecosystem. YOLOv5 is heavily optimized for fast [training efficiency](https://docs.ultralytics.com/guides/model-training-tips/) and low-memory environments, making it incredibly stable for edge deployments. +Its crowning achievement, however, is its integration into the broader Ultralytics ecosystem. YOLOv5 is heavily optimized for fast [training efficiency](https://docs.ultralytics.com/guides/model-training-tips) and low-memory environments, making it incredibly stable for edge deployments. [Learn more about YOLOv5](https://platform.ultralytics.com/ultralytics/yolov5){ .md-button } @@ -39,7 +39,7 @@ Its crowning achievement, however, is its integration into the broader Ultralyti ## Performance Analysis: Speed vs. Accuracy -When designing a computer vision pipeline, developers must weigh the trade-offs between precision and latency. The following table illustrates the performance differences on the standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). +When designing a computer vision pipeline, developers must weigh the trade-offs between precision and latency. The following table illustrates the performance differences on the standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -59,19 +59,19 @@ When designing a computer vision pipeline, developers must weigh the trade-offs YOLOv9 establishes absolute dominance in raw precision. The **YOLOv9e** pushes the boundaries of mAP to 55.6%, utilizing its GELAN layers to preserve fine-grained details. This makes it an exceptional choice for [medical imaging](https://www.ultralytics.com/solutions/ai-in-healthcare) or scenarios demanding rigorous accuracy on small objects. -Conversely, **YOLOv5** shines in its raw deployment speed and hardware flexibility. The YOLOv5n (Nano) is famously lightweight, executing inferences in just 1.12ms on a T4 GPU via [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). If you are deploying to constrained IoT devices, mobile phones, or [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/), the memory footprint of YOLOv5 makes it extraordinarily reliable. +Conversely, **YOLOv5** shines in its raw deployment speed and hardware flexibility. The YOLOv5n (Nano) is famously lightweight, executing inferences in just 1.12ms on a T4 GPU via [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). If you are deploying to constrained IoT devices, mobile phones, or [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi), the memory footprint of YOLOv5 makes it extraordinarily reliable. ## The Ultralytics Ecosystem Advantage -A major consideration when selecting a model is the surrounding software ecosystem. While YOLOv9 provides top-tier research benchmarks, utilizing both models through the modern [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) bridges the gap, offering developers a unified and streamlined experience. +A major consideration when selecting a model is the surrounding software ecosystem. While YOLOv9 provides top-tier research benchmarks, utilizing both models through the modern [Ultralytics Python API](https://docs.ultralytics.com/usage/python) bridges the gap, offering developers a unified and streamlined experience. ### Ease of Use and Exporting -Ultralytics abstracts complex engineering hurdles. Features like automatic [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation/) and [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/) are handled out of the box. Moving models to production is equally trivial, with built-in export commands to convert models into [ONNX](https://docs.ultralytics.com/integrations/onnx/), [OpenVINO](https://docs.ultralytics.com/integrations/openvino/), or [TFLite](https://docs.ultralytics.com/integrations/tflite/) formats. +Ultralytics abstracts complex engineering hurdles. Features like automatic [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation) and [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning) are handled out of the box. Moving models to production is equally trivial, with built-in export commands to convert models into [ONNX](https://docs.ultralytics.com/integrations/onnx), [OpenVINO](https://docs.ultralytics.com/integrations/openvino), or [TFLite](https://docs.ultralytics.com/integrations/tflite) formats. ### Task Versatility -While both models excel at [object detection](https://docs.ultralytics.com/tasks/detect/), modern Ultralytics models are built to tackle a variety of computer vision challenges. The broader framework provides native support for [image classification](https://docs.ultralytics.com/tasks/classify/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), allowing developers to solve multiple vision problems without switching codebases. +While both models excel at [object detection](https://docs.ultralytics.com/tasks/detect), modern Ultralytics models are built to tackle a variety of computer vision challenges. The broader framework provides native support for [image classification](https://docs.ultralytics.com/tasks/classify), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb), allowing developers to solve multiple vision problems without switching codebases. ## Use Cases and Recommendations @@ -91,15 +91,15 @@ YOLOv5 is recommended for: - **Proven Production Systems:** Existing deployments where YOLOv5's long track record of stability, extensive documentation, and massive community support are valued. - **Resource-Constrained Training:** Environments with limited GPU resources where YOLOv5's efficient training pipeline and lower memory requirements are advantageous. -- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), [CoreML](https://docs.ultralytics.com/integrations/coreml/), and [TFLite](https://docs.ultralytics.com/integrations/tflite/). +- **Extensive Export Format Support:** Projects requiring deployment across many formats including [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), [CoreML](https://docs.ultralytics.com/integrations/coreml), and [TFLite](https://docs.ultralytics.com/integrations/tflite). ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Implementation Example diff --git a/docs/en/compare/yolov9-vs-yolov6.md b/docs/en/compare/yolov9-vs-yolov6.md index 5140bbddb98..f8868d61d12 100644 --- a/docs/en/compare/yolov9-vs-yolov6.md +++ b/docs/en/compare/yolov9-vs-yolov6.md @@ -28,7 +28,7 @@ Introduced in early 2024, YOLOv9 tackles one of the most persistent challenges i YOLOv9 introduces the concept of Programmable Gradient Information (PGI) alongside the Generalized Efficient Layer Aggregation Network (GELAN). PGI addresses the information bottleneck by providing auxiliary supervision that ensures the main network learns robust, reliable features without adding inference overhead. Meanwhile, GELAN optimizes parameter utilization, allowing the model to achieve state-of-the-art [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map) while keeping the computational cost manageable. This makes it an exceptional choice for [medical image analysis](https://www.ultralytics.com/glossary/medical-image-analysis) or detecting extremely small objects where feature fidelity is critical. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ## YOLOv6-3.0 Overview: Built for Industrial Scale @@ -41,13 +41,13 @@ Developed by Meituan, YOLOv6-3.0 (also referred to as v3.0) is engineered from t ### Architecture and Methodologies -YOLOv6-3.0 distinguishes itself through its RepOptimizer and Anchor-Aided Training (AAT) strategies. The model utilizes a hardware-aware neural network design inspired by RepVGG, which allows it to run exceptionally fast on GPUs during inference by fusing layers. The 3.0 update further refined the architecture by introducing a Bi-directional Concatenation (BiC) module to improve localization accuracy. Because it is highly optimized for deployment formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino/), YOLOv6-3.0 is frequently adopted in logistics, [manufacturing automation](https://www.ultralytics.com/blog/manufacturing-automation), and high-throughput server environments. +YOLOv6-3.0 distinguishes itself through its RepOptimizer and Anchor-Aided Training (AAT) strategies. The model utilizes a hardware-aware neural network design inspired by RepVGG, which allows it to run exceptionally fast on GPUs during inference by fusing layers. The 3.0 update further refined the architecture by introducing a Bi-directional Concatenation (BiC) module to improve localization accuracy. Because it is highly optimized for deployment formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino), YOLOv6-3.0 is frequently adopted in logistics, [manufacturing automation](https://www.ultralytics.com/blog/manufacturing-automation), and high-throughput server environments. -[Learn more about YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6-3.0](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Performance Comparison -When evaluating these models on the standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/), we can observe distinct trade-offs between accuracy and raw inference speed. +When evaluating these models on the standard [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco), we can observe distinct trade-offs between accuracy and raw inference speed. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ----------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -68,17 +68,17 @@ While YOLOv6-3.0n takes the crown for raw speed on T4 hardware (1.17ms), YOLOv9t !!! tip "Future-Proof Your Project with YOLO26" - If you are starting a new computer vision initiative, we highly recommend utilizing **[YOLO26](https://docs.ultralytics.com/models/yolo26/)**. Released in 2026, it features a native **End-to-End NMS-Free Design** that completely eliminates post-processing latency, unlocking up to **43% Faster CPU Inference**. + If you are starting a new computer vision initiative, we highly recommend utilizing **[YOLO26](https://docs.ultralytics.com/models/yolo26)**. Released in 2026, it features a native **End-to-End NMS-Free Design** that completely eliminates post-processing latency, unlocking up to **43% Faster CPU Inference**. ## The Ultralytics Ecosystem Advantage -Regardless of which model's architectural philosophy appeals to you, implementing them natively through the [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) provides a superior developer experience. +Regardless of which model's architectural philosophy appeals to you, implementing them natively through the [Ultralytics Python API](https://docs.ultralytics.com/usage/python) provides a superior developer experience. ### Ease of Use and Training Efficiency Training complex deep learning models traditionally requires massive boilerplate code. The [Ultralytics Platform](https://platform.ultralytics.com) abstracts these complexities. Whether you are fine-tuning YOLOv9 for [defect detection](https://www.ultralytics.com/blog/how-vision-ai-enhances-defect-detection-on-production-lines) or exporting YOLOv6 for mobile applications, the workflow remains remarkably consistent. -Furthermore, Ultralytics architectures generally boast lower [CUDA memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics/) during training compared to bulky transformer-based models. This allows developers to use larger batch sizes on consumer-grade GPUs, vastly improving training efficiency. +Furthermore, Ultralytics architectures generally boast lower [CUDA memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics) during training compared to bulky transformer-based models. This allows developers to use larger batch sizes on consumer-grade GPUs, vastly improving training efficiency. ```python from ultralytics import YOLO @@ -96,7 +96,7 @@ model.export(format="engine", half=True) ### Unmatched Versatility Across Vision Tasks -While YOLOv6-3.0 is heavily optimized for fast bounding box generation, modern computer vision projects often require a multi-task approach. Ultralytics models are celebrated for their extreme versatility. With tools like [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) and the newer YOLO26, a single framework seamlessly handles [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +While YOLOv6-3.0 is heavily optimized for fast bounding box generation, modern computer vision projects often require a multi-task approach. Ultralytics models are celebrated for their extreme versatility. With tools like [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) and the newer YOLO26, a single framework seamlessly handles [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb). ## Introducing YOLO26: The New Standard diff --git a/docs/en/compare/yolov9-vs-yolov7.md b/docs/en/compare/yolov9-vs-yolov7.md index e268a5af712..17f1ee6c995 100644 --- a/docs/en/compare/yolov9-vs-yolov7.md +++ b/docs/en/compare/yolov9-vs-yolov7.md @@ -8,7 +8,7 @@ keywords: YOLOv9, YOLOv7, object detection, AI models, technical comparison, neu The evolution of real-time [object detection](https://en.wikipedia.org/wiki/Object_detection) has been driven by a continuous quest to balance computational efficiency with high accuracy. Two landmark architectures in this journey are YOLOv9 and YOLOv7, both developed by researchers at the Institute of Information Science, Academia Sinica in Taiwan. While YOLOv7 introduced revolutionary trainable bag-of-freebies, the newer YOLOv9 tackles deep learning information bottlenecks head-on. -This comprehensive technical comparison explores the architectural differences, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics/), and ideal deployment scenarios for both models, helping ML engineers and researchers choose the right tool for their computer vision pipelines. +This comprehensive technical comparison explores the architectural differences, [performance metrics](https://docs.ultralytics.com/guides/yolo-performance-metrics), and ideal deployment scenarios for both models, helping ML engineers and researchers choose the right tool for their computer vision pipelines. @@ -32,7 +32,7 @@ When comparing these models, raw performance and efficiency are critical factors !!! tip "Performance Balance" - Notice how YOLOv9c achieves roughly the same accuracy (53.0 mAP) as YOLOv7x (53.1 mAP) while utilizing significantly fewer parameters (25.3M vs 71.3M) and FLOPs. This demonstrates the [Performance Balance](https://docs.ultralytics.com/guides/yolo-performance-metrics/) improvements in modern architectures. + Notice how YOLOv9c achieves roughly the same accuracy (53.0 mAP) as YOLOv7x (53.1 mAP) while utilizing significantly fewer parameters (25.3M vs 71.3M) and FLOPs. This demonstrates the [Performance Balance](https://docs.ultralytics.com/guides/yolo-performance-metrics) improvements in modern architectures. ## YOLOv9: Solving the Information Bottleneck @@ -51,7 +51,7 @@ YOLOv9 introduces the Generalized Efficient Layer Aggregation Network (GELAN) an The main strength of YOLOv9 is its ability to extract subtle features without immense computational overhead, making it incredibly capable for tasks requiring high feature fidelity, like [medical image analysis](https://www.ultralytics.com/glossary/medical-image-analysis). However, the complex PGI structure during training can make custom architectural modifications more challenging for beginners compared to more unified frameworks. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ## YOLOv7: The Bag-of-Freebies Pioneer @@ -68,9 +68,9 @@ YOLOv7's core contribution is the Extended Efficient Layer Aggregation Network ( ### Strengths and Limitations -YOLOv7 is highly optimized for real-time edge processing and remains a staple in legacy systems and older [CUDA environments](https://developer.nvidia.com/cuda/toolkit). Its primary limitation today is its larger parameter size compared to newer models. As shown in the performance table, achieving top-tier accuracy requires the heavy YOLOv7x model, which demands substantially more [GPU memory](https://docs.ultralytics.com/guides/yolo-performance-metrics/) than equivalent modern architectures. +YOLOv7 is highly optimized for real-time edge processing and remains a staple in legacy systems and older [CUDA environments](https://developer.nvidia.com/cuda/toolkit). Its primary limitation today is its larger parameter size compared to newer models. As shown in the performance table, achieving top-tier accuracy requires the heavy YOLOv7x model, which demands substantially more [GPU memory](https://docs.ultralytics.com/guides/yolo-performance-metrics) than equivalent modern architectures. -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ## The Ultralytics Advantage: Streamlined Deployment @@ -86,9 +86,9 @@ If you are starting a new computer vision project, we highly recommend exploring - **Up to 43% Faster CPU Inference:** Optimized for [edge computing](https://www.ultralytics.com/glossary/edge-computing) environments, ensuring your application runs smoothly even without dedicated GPUs. - **MuSGD Optimizer:** A hybrid optimizer inspired by LLM training, delivering highly stable convergence and reducing training time. - **DFL Removal:** Simplified model export by removing Distribution Focal Loss, enhancing compatibility with low-power mobile devices. -- **ProgLoss + STAL:** Drastically improves performance on small object detection, making it the premier choice for [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and surveillance. +- **ProgLoss + STAL:** Drastically improves performance on small object detection, making it the premier choice for [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and surveillance. -Other popular alternatives within the ecosystem include [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) and [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), both of which offer massive versatility across tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [pose estimation](https://docs.ultralytics.com/tasks/pose/). +Other popular alternatives within the ecosystem include [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) and [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11), both of which offer massive versatility across tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [pose estimation](https://docs.ultralytics.com/tasks/pose). ### Implementation Example diff --git a/docs/en/compare/yolov9-vs-yolov8.md b/docs/en/compare/yolov9-vs-yolov8.md index 9ff25a1f04a..7e1a9dcd8b6 100644 --- a/docs/en/compare/yolov9-vs-yolov8.md +++ b/docs/en/compare/yolov9-vs-yolov8.md @@ -6,7 +6,7 @@ keywords: YOLOv9, YOLOv8, object detection, computer vision, YOLO comparison, de # YOLOv9 vs. YOLOv8: A Technical Deep Dive into Modern Object Detection -The landscape of real-time computer vision has evolved remarkably over the last few years, with each new model pushing the theoretical boundaries of what is possible on edge devices and cloud servers alike. When comparing the newer [YOLOv9 architecture](https://docs.ultralytics.com/models/yolov9/) to the highly popular [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) framework, developers are often faced with a choice between cutting-edge theoretical gradient paths and a heavily battle-tested, production-ready ecosystem. +The landscape of real-time computer vision has evolved remarkably over the last few years, with each new model pushing the theoretical boundaries of what is possible on edge devices and cloud servers alike. When comparing the newer [YOLOv9 architecture](https://docs.ultralytics.com/models/yolov9) to the highly popular [Ultralytics YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8) framework, developers are often faced with a choice between cutting-edge theoretical gradient paths and a heavily battle-tested, production-ready ecosystem. This comprehensive guide contrasts these two heavyweights, analyzing their architectural innovations, performance metrics, and ideal deployment scenarios to help you choose the right model for your next artificial intelligence project. @@ -22,7 +22,7 @@ Understanding the lineage of these models provides essential context for their r **YOLOv9** Authored by Chien-Yao Wang and Hong-Yuan Mark Liao at the Institute of Information Science, Academia Sinica, Taiwan, YOLOv9 was released on February 21, 2024. The core research focuses on solving the information bottleneck in deep neural networks. You can explore the original [YOLOv9 research paper](https://arxiv.org/abs/2402.13616) on Arxiv or view the source code in the [official YOLOv9 GitHub repository](https://github.com/WongKinYiu/yolov9). -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } **Ultralytics YOLOv8** Developed by Glenn Jocher, Ayush Chaurasia, and Jing Qiu at Ultralytics, YOLOv8 launched on January 10, 2023. It established itself as an industry standard for versatility, offering a unified API for a massive variety of vision tasks. The source code is maintained within the main [Ultralytics GitHub repository](https://github.com/ultralytics/ultralytics), ensuring continuous updates and long-term stability. @@ -37,7 +37,7 @@ The defining feature of YOLOv9 is its introduction of **Programmable Gradient In ### YOLOv8: The Versatile Workhorse -YOLOv8 introduced a streamlined anchor-free detection mechanism, which reduces the number of box predictions and speeds up Non-Maximum Suppression (NMS) during post-processing. Its C2f module (Cross-Stage Partial Bottleneck with two convolutions) improves gradient flow across the network compared to older models. More importantly, YOLOv8 was designed with **Versatility** in mind, natively supporting [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb/) extraction out of the box. +YOLOv8 introduced a streamlined anchor-free detection mechanism, which reduces the number of box predictions and speeds up Non-Maximum Suppression (NMS) during post-processing. Its C2f module (Cross-Stage Partial Bottleneck with two convolutions) improves gradient flow across the network compared to older models. More importantly, YOLOv8 was designed with **Versatility** in mind, natively supporting [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb) extraction out of the box. !!! tip "Ecosystem Integration" @@ -67,7 +67,7 @@ When analyzing the metrics, YOLOv9 demonstrates a remarkable parameter-to-accura A major consideration when choosing an architecture is the **Ease of Use** and the surrounding software ecosystem. Managing dependencies, writing custom data loaders, and handling complex export scripts can stall development. The integrated Ultralytics ecosystem abstracts these complexities away. -Whether you choose YOLOv8 or YOLOv9 (which is fully supported within the Ultralytics library), you benefit from a unified API, automatic [data augmentation techniques](https://docs.ultralytics.com/guides/yolo-data-augmentation/), and streamlined [ONNX format](https://onnx.ai/) exporting. Furthermore, Ultralytics architectures generally feature highly optimized **Training Efficiency**, avoiding the massive CUDA memory bloat commonly associated with large transformer-based models. +Whether you choose YOLOv8 or YOLOv9 (which is fully supported within the Ultralytics library), you benefit from a unified API, automatic [data augmentation techniques](https://docs.ultralytics.com/guides/yolo-data-augmentation), and streamlined [ONNX format](https://onnx.ai/) exporting. Furthermore, Ultralytics architectures generally feature highly optimized **Training Efficiency**, avoiding the massive CUDA memory bloat commonly associated with large transformer-based models. ### Training Code Example @@ -105,17 +105,17 @@ YOLOv9 is a strong choice for: YOLOv8 is recommended for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Looking Forward: The Arrival of YOLO26 @@ -129,7 +129,7 @@ Furthermore, YOLO26 utilizes the groundbreaking **MuSGD Optimizer**, a hybrid of !!! info "Alternative Architectures" - Depending on your hardware constraints, you may also be interested in comparing these models with [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) for balanced general-purpose tasks, or exploring transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) for specialized high-fidelity research. + Depending on your hardware constraints, you may also be interested in comparing these models with [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) for balanced general-purpose tasks, or exploring transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) for specialized high-fidelity research. ## Real-World Applications and Use Cases @@ -138,4 +138,4 @@ The choice between YOLOv8 and YOLOv9 largely depends on your project constraints - **Healthcare and Medical Imaging:** When every pixel counts, such as in [tumor detection systems](https://www.ultralytics.com/solutions/ai-in-healthcare), YOLOv9's GELAN architecture preserves fine-grained details exceptionally well, reducing false negatives in critical diagnoses. - **Retail and Inventory Analytics:** For [smart supermarket systems](https://www.ultralytics.com/solutions/ai-in-retail) tracking densely packed shelves, YOLOv9 provides the necessary mAP to separate overlapping items reliably. - **Smart Cities and Traffic Monitoring:** In fast-paced [logistics and traffic management](https://www.ultralytics.com/solutions/ai-in-logistics), the ultra-low latency and proven robustness of YOLOv8 make it ideal for tracking vehicles across multiple camera streams simultaneously. -- **Edge Deployments:** If you are deploying to constrained devices like a Raspberry Pi or [mobile hardware](https://docs.ultralytics.com/integrations/tflite/), the highly optimized C2f blocks of YOLOv8 (and the CPU optimizations of YOLO26) provide a much smoother, battery-friendly inference pipeline. +- **Edge Deployments:** If you are deploying to constrained devices like a Raspberry Pi or [mobile hardware](https://docs.ultralytics.com/integrations/tflite), the highly optimized C2f blocks of YOLOv8 (and the CPU optimizations of YOLO26) provide a much smoother, battery-friendly inference pipeline. diff --git a/docs/en/compare/yolov9-vs-yolox.md b/docs/en/compare/yolov9-vs-yolox.md index 88718e37eed..09461b1a005 100644 --- a/docs/en/compare/yolov9-vs-yolox.md +++ b/docs/en/compare/yolov9-vs-yolox.md @@ -6,7 +6,7 @@ keywords: YOLOv9, YOLOX, object detection, model comparison, computer vision, YO # YOLOv9 vs YOLOX: A Technical Deep Dive into Modern Object Detection -The field of computer vision has witnessed a rapid evolution in real-time object detection architectures. This guide provides a comprehensive comparison between **[YOLOv9](https://docs.ultralytics.com/models/yolov9/)** and **YOLOX**, analyzing their architectural innovations, performance metrics, and training methodologies. Whether you are building smart applications for [AI in manufacturing](https://www.ultralytics.com/solutions/ai-in-manufacturing) or exploring [predictive modeling](https://www.ultralytics.com/glossary/predictive-modeling), understanding these models will help you make informed decisions for your next deployment. +The field of computer vision has witnessed a rapid evolution in real-time object detection architectures. This guide provides a comprehensive comparison between **[YOLOv9](https://docs.ultralytics.com/models/yolov9)** and **YOLOX**, analyzing their architectural innovations, performance metrics, and training methodologies. Whether you are building smart applications for [AI in manufacturing](https://www.ultralytics.com/solutions/ai-in-manufacturing) or exploring [predictive modeling](https://www.ultralytics.com/glossary/predictive-modeling), understanding these models will help you make informed decisions for your next deployment. @@ -27,7 +27,7 @@ YOLOv9 introduced a paradigm shift by addressing the information bottleneck prob By retaining crucial feature data during the feed-forward process, YOLOv9 ensures that the gradients used to update weights during backpropagation remain accurate. This architecture excels at [feature extraction](https://www.ultralytics.com/glossary/feature-extraction), making it highly capable of detecting small objects in complex environments, such as those found in [aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision) and detailed medical scans. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } ### YOLOX: Bridging Research and Industry @@ -90,11 +90,11 @@ metrics = model.val() model.export(format="engine") ``` -With built-in support for multiple tasks, including [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/), you can rapidly pivot your computer vision solutions without changing your entire codebase. +With built-in support for multiple tasks, including [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), and [pose estimation](https://docs.ultralytics.com/tasks/pose), you can rapidly pivot your computer vision solutions without changing your entire codebase. !!! tip "Seamless Exporting" - Deploying to the edge? Ultralytics makes it simple to export your trained models to highly optimized formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), and OpenVINO with just a single command. + Deploying to the edge? Ultralytics makes it simple to export your trained models to highly optimized formats like [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), and OpenVINO with just a single command. ## Real-World Applications @@ -134,11 +134,11 @@ YOLOX is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Future: Enter YOLO26 @@ -148,4 +148,4 @@ YOLO26 completely revitalizes the deployment pipeline with a native **End-to-End Furthermore, YOLO26 incorporates the groundbreaking **MuSGD Optimizer**, a hybrid of SGD and Muon that borrows innovations from LLM training to provide incredibly stable and rapid convergence. By removing Distribution Focal Loss (DFL), YOLO26 achieves up to **43% faster CPU inference** compared to its predecessors, making it the absolute best choice for edge devices and enterprise deployments. With notable improvements in small-object recognition via ProgLoss and STAL, YOLO26 effectively supersedes both YOLOX and YOLOv9. -For engineers exploring modern architectures, we also recommend checking out [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) as powerful alternatives within the Ultralytics suite. Ensure your project is future-proofed by leveraging the unparalleled performance of the latest models on the Ultralytics Platform. +For engineers exploring modern architectures, we also recommend checking out [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) and [RT-DETR](https://docs.ultralytics.com/models/rtdetr) as powerful alternatives within the Ultralytics suite. Ensure your project is future-proofed by leveraging the unparalleled performance of the latest models on the Ultralytics Platform. diff --git a/docs/en/compare/yolox-vs-damo-yolo.md b/docs/en/compare/yolox-vs-damo-yolo.md index c49d5edceba..9e3a442262c 100644 --- a/docs/en/compare/yolox-vs-damo-yolo.md +++ b/docs/en/compare/yolox-vs-damo-yolo.md @@ -39,7 +39,7 @@ Developed by Xianzhe Xu and a team of researchers at the [Alibaba Group](https:/ DAMO-YOLO's strategy was built on automating the design of efficient structures: -- **MAE-NAS Backbones:** Utilizing a Multi-Objective Evolutionary algorithm, DAMO-YOLO discovered highly efficient backbones customized for specific latency budgets, particularly when exported to frameworks like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). +- **MAE-NAS Backbones:** Utilizing a Multi-Objective Evolutionary algorithm, DAMO-YOLO discovered highly efficient backbones customized for specific latency budgets, particularly when exported to frameworks like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). - **Efficient RepGFPN:** A heavy-neck design that significantly enhances feature fusion across different spatial resolutions, which is highly beneficial for [aerial imagery analysis](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision) and detecting objects at varying scales. - **ZeroHead:** A simplified prediction head that trims computational redundancy without sacrificing the model's overall mean Average Precision (mAP). - **AlignedOTA and Distillation:** Incorporates advanced label assignment and teacher-student knowledge distillation to squeeze maximum performance out of smaller student models. @@ -88,15 +88,15 @@ DAMO-YOLO is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Introducing YOLO26 -While YOLOX and DAMO-YOLO represent important historical milestones, modern developers require a solution that pairs state-of-the-art accuracy with unparalleled ease of use. This is where [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) transforms the landscape. Released in January 2026, YOLO26 builds upon the legacy of [NMS-free models](https://docs.ultralytics.com/models/yolov10/) to deliver the ultimate balance of speed, accuracy, and developer experience. +While YOLOX and DAMO-YOLO represent important historical milestones, modern developers require a solution that pairs state-of-the-art accuracy with unparalleled ease of use. This is where [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) transforms the landscape. Released in January 2026, YOLO26 builds upon the legacy of [NMS-free models](https://docs.ultralytics.com/models/yolov10) to deliver the ultimate balance of speed, accuracy, and developer experience. ### Why Choose YOLO26? @@ -107,11 +107,11 @@ The integrated Ultralytics ecosystem outshines fragmented academic repositories - **MuSGD Optimizer:** YOLO26 borrows LLM training innovations with its hybrid SGD and Muon optimizer, ensuring rock-solid training stability and ultra-fast convergence. - **Up to 43% Faster CPU Inference:** Thanks to deep structural optimizations, YOLO26 runs blazingly fast on CPUs without needing expensive GPU hardware. - **Advanced Loss Functions:** The integration of ProgLoss + STAL provides massive improvements in small-object recognition, making it ideal for tasks like [drone inspections](https://www.ultralytics.com/solutions/ai-in-agriculture) and IoT monitoring. -- **Versatility:** Unlike DAMO-YOLO, which is strictly a detector, YOLO26 natively supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), [Image Classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) tasks in a single, unified framework. +- **Versatility:** Unlike DAMO-YOLO, which is strictly a detector, YOLO26 natively supports [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), [Image Classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) tasks in a single, unified framework. !!! tip "Start Building Instantly" - With the [Ultralytics Python API](https://docs.ultralytics.com/usage/python/), you don't need to manually configure complex distillation pipelines or write hundreds of lines of C++ code to deploy your model. + With the [Ultralytics Python API](https://docs.ultralytics.com/usage/python), you don't need to manually configure complex distillation pipelines or write hundreds of lines of C++ code to deploy your model. ```python from ultralytics import YOLO @@ -137,7 +137,7 @@ The computer vision ecosystem is vast. Depending on your specific constraints, y - **[YOLO11](https://platform.ultralytics.com/ultralytics/yolo11):** The highly capable predecessor to YOLO26, known for its robustness in [retail analytics](https://www.ultralytics.com/solutions/ai-in-retail) and [manufacturing quality control](https://www.ultralytics.com/solutions/ai-in-manufacturing). - **[YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8):** A legendary, highly stable anchor-free model that popularized widespread edge deployment. -- **[RT-DETR](https://docs.ultralytics.com/models/rtdetr/):** A Real-Time DEtection TRansformer developed by Baidu, offering an excellent alternative for tasks that benefit heavily from global attention mechanisms, albeit at the cost of higher training memory requirements. +- **[RT-DETR](https://docs.ultralytics.com/models/rtdetr):** A Real-Time DEtection TRansformer developed by Baidu, offering an excellent alternative for tasks that benefit heavily from global attention mechanisms, albeit at the cost of higher training memory requirements. ## Conclusion diff --git a/docs/en/compare/yolox-vs-efficientdet.md b/docs/en/compare/yolox-vs-efficientdet.md index 0e1b9ece1e4..77d846ac927 100644 --- a/docs/en/compare/yolox-vs-efficientdet.md +++ b/docs/en/compare/yolox-vs-efficientdet.md @@ -6,7 +6,7 @@ keywords: YOLOX, EfficientDet, object detection, model comparison, deep learning # YOLOX vs. EfficientDet: Evaluating Anchor-Free and Scalable Object Detection -The evolution of [object detection](https://docs.ultralytics.com/tasks/detect/) has been driven by the constant pursuit of balancing speed, accuracy, and computational efficiency. Two landmark models that significantly influenced this trajectory are YOLOX and EfficientDet. While YOLOX introduced a highly optimized anchor-free design to the YOLO family, EfficientDet focused on a scalable architecture utilizing compound scaling and BiFPN. This guide provides a detailed technical comparison of their architectures, performance metrics, and training methodologies, while also introducing modern alternatives like the cutting-edge [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) model. +The evolution of [object detection](https://docs.ultralytics.com/tasks/detect) has been driven by the constant pursuit of balancing speed, accuracy, and computational efficiency. Two landmark models that significantly influenced this trajectory are YOLOX and EfficientDet. While YOLOX introduced a highly optimized anchor-free design to the YOLO family, EfficientDet focused on a scalable architecture utilizing compound scaling and BiFPN. This guide provides a detailed technical comparison of their architectures, performance metrics, and training methodologies, while also introducing modern alternatives like the cutting-edge [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) model. ## Model Origins and Technical Details @@ -106,11 +106,11 @@ EfficientDet is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Modern Alternative: Ultralytics YOLO26 @@ -122,15 +122,15 @@ YOLO26 offers a well-maintained ecosystem and a massive leap forward in both spe - **End-to-End NMS-Free Design:** YOLO26 eliminates the need for [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing. This natively end-to-end approach, pioneered in earlier generations, simplifies the export process and slashes deployment latency. - **Up to 43% Faster CPU Inference:** Thanks to deep architectural optimizations and the removal of Distribution Focal Loss (DFL), YOLO26 is remarkably fast on edge devices lacking discrete GPUs, far outpacing the heavy EfficientDet variants. -- **MuSGD Optimizer:** Bringing [Large Language Model (LLM)](https://www.ultralytics.com/glossary/large-language-model-llm) innovations to vision, YOLO26 utilizes the MuSGD optimizer (a hybrid of SGD and Muon) for highly stable training and rapid convergence, resulting in excellent [training efficiency](https://docs.ultralytics.com/guides/model-training-tips/). +- **MuSGD Optimizer:** Bringing [Large Language Model (LLM)](https://www.ultralytics.com/glossary/large-language-model-llm) innovations to vision, YOLO26 utilizes the MuSGD optimizer (a hybrid of SGD and Muon) for highly stable training and rapid convergence, resulting in excellent [training efficiency](https://docs.ultralytics.com/guides/model-training-tips). - **ProgLoss + STAL:** These advanced loss functions yield notable improvements in small-object recognition, which is critical for use cases like [drone operations](https://www.ultralytics.com/blog/computer-vision-applications-ai-drone-uav-operations) and aerial imagery analysis. -- **Unmatched Versatility:** Unlike YOLOX, which is strictly an object detector, YOLO26 natively supports a wide array of tasks including [instance segmentation](https://docs.ultralytics.com/tasks/segment/), image classification, [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. +- **Unmatched Versatility:** Unlike YOLOX, which is strictly an object detector, YOLO26 natively supports a wide array of tasks including [instance segmentation](https://docs.ultralytics.com/tasks/segment), image classification, [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } ### Ease of Use with the Ultralytics API -One of the most significant advantages of Ultralytics models is the streamlined user experience. Training and deploying a YOLO26 model requires drastically lower [memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics/) than complex transformer models and involves just a few lines of Python code: +One of the most significant advantages of Ultralytics models is the streamlined user experience. Training and deploying a YOLO26 model requires drastically lower [memory requirements](https://docs.ultralytics.com/guides/yolo-performance-metrics) than complex transformer models and involves just a few lines of Python code: ```python from ultralytics import YOLO @@ -145,7 +145,7 @@ results = model.train(data="coco8.yaml", epochs=100, imgsz=640) model.export(format="engine", dynamic=True) ``` -For users who prefer visual interfaces, the [Ultralytics Platform](https://docs.ultralytics.com/platform/) provides powerful tools for dataset annotation, hyperparameter tuning, and seamless deployment. +For users who prefer visual interfaces, the [Ultralytics Platform](https://docs.ultralytics.com/platform) provides powerful tools for dataset annotation, hyperparameter tuning, and seamless deployment. ## Real-World Use Cases @@ -161,6 +161,6 @@ YOLOX is suitable for applications requiring a balance of speed and accuracy wit ### Why YOLO26 is the Superior Choice -For almost all modern applications, YOLO26 provides the best solution. Its NMS-free design ensures deterministic latency, making it the perfect candidate for autonomous driving, rapid [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/), and smart city deployments. Furthermore, the robust community support and frequent updates from Ultralytics ensure that developers are never left dealing with deprecated dependencies. +For almost all modern applications, YOLO26 provides the best solution. Its NMS-free design ensures deterministic latency, making it the perfect candidate for autonomous driving, rapid [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system), and smart city deployments. Furthermore, the robust community support and frequent updates from Ultralytics ensure that developers are never left dealing with deprecated dependencies. -Developers exploring advanced computer vision should also look into other versatile architectures within the Ultralytics ecosystem, such as [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) for stable legacy deployments or specialized models like [FastSAM](https://docs.ultralytics.com/models/fast-sam/) for prompt-based segmentation tasks. Utilizing the full suite of Ultralytics tools guarantees a future-proof, highly optimized vision AI pipeline. +Developers exploring advanced computer vision should also look into other versatile architectures within the Ultralytics ecosystem, such as [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) for stable legacy deployments or specialized models like [FastSAM](https://docs.ultralytics.com/models/fast-sam) for prompt-based segmentation tasks. Utilizing the full suite of Ultralytics tools guarantees a future-proof, highly optimized vision AI pipeline. diff --git a/docs/en/compare/yolox-vs-pp-yoloe.md b/docs/en/compare/yolox-vs-pp-yoloe.md index ef637118cab..5c243d717a4 100644 --- a/docs/en/compare/yolox-vs-pp-yoloe.md +++ b/docs/en/compare/yolox-vs-pp-yoloe.md @@ -29,13 +29,13 @@ YOLOX integrates a decoupled head, separating classification and regression task Introduced by the PaddlePaddle Authors at [Baidu](https://www.baidu.com/) on April 2, 2022, PP-YOLOE+ represents a highly optimized evolution of the PP-YOLO series. Detailed in their [Arxiv publication](https://arxiv.org/abs/2203.16250), PP-YOLOE+ is deeply integrated into the Baidu ecosystem and requires the PaddlePaddle framework. The model's configurations can be found in the [PaddleDetection GitHub repository](https://github.com/PaddlePaddle/PaddleDetection/). -PP-YOLOE+ relies on a powerful CSPRepResNet backbone and utilizes an Efficient Task-aligned head (ET-head) alongside Task Alignment Learning (TAL). This architecture achieves outstanding [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map) on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/), making it a formidable choice for industrial defect detection and heavy server-side processing where accuracy is prioritized over minimal dependencies. +PP-YOLOE+ relies on a powerful CSPRepResNet backbone and utilizes an Efficient Task-aligned head (ET-head) alongside Task Alignment Learning (TAL). This architecture achieves outstanding [mean Average Precision (mAP)](https://www.ultralytics.com/glossary/mean-average-precision-map) on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco), making it a formidable choice for industrial defect detection and heavy server-side processing where accuracy is prioritized over minimal dependencies. [Learn more about PP-YOLOE+](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.8.1/configs/ppyoloe/README.md){ .md-button } ## Performance Benchmarks -Understanding how these models perform across different scales is essential for deployment. The table below outlines key metrics, including [mAP](https://docs.ultralytics.com/guides/yolo-performance-metrics/) and inference speeds when exported to [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). +Understanding how these models perform across different scales is essential for deployment. The table below outlines key metrics, including [mAP](https://docs.ultralytics.com/guides/yolo-performance-metrics) and inference speeds when exported to [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -78,24 +78,24 @@ PP-YOLOE+ is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: Introducing YOLO26 -While both YOLOX and PP-YOLOE+ offer distinct advantages, the rapid evolution of AI demands tools that combine state-of-the-art accuracy with unparalleled ease of use. This is where [Ultralytics](https://www.ultralytics.com/) models, specifically the recently released [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/), outshine legacy research repositories. +While both YOLOX and PP-YOLOE+ offer distinct advantages, the rapid evolution of AI demands tools that combine state-of-the-art accuracy with unparalleled ease of use. This is where [Ultralytics](https://www.ultralytics.com/) models, specifically the recently released [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26), outshine legacy research repositories. -Released in January 2026, YOLO26 establishes a new standard for modern [object detection](https://docs.ultralytics.com/tasks/detect/) and beyond, offering a developer experience that is simply unmatched by competing frameworks. +Released in January 2026, YOLO26 establishes a new standard for modern [object detection](https://docs.ultralytics.com/tasks/detect) and beyond, offering a developer experience that is simply unmatched by competing frameworks. ### Why Developers Choose YOLO26 -1. **End-to-End NMS-Free Design:** Building on concepts pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10/), YOLO26 is natively end-to-end. By entirely removing Non-Maximum Suppression (NMS) post-processing, it ensures highly consistent latency and dramatically simplifies export pipelines for edge environments. -2. **Next-Generation Optimization:** Training stability is revolutionized by the **MuSGD Optimizer**, a hybrid of SGD and Muon (inspired by LLM methodologies like Moonshot AI's Kimi K2). This guarantees faster convergence. Furthermore, YOLO26 utilizes **ProgLoss + STAL** to drastically improve small-object recognition, a crucial feature for applications involving [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and robotics. +1. **End-to-End NMS-Free Design:** Building on concepts pioneered in [YOLOv10](https://docs.ultralytics.com/models/yolov10), YOLO26 is natively end-to-end. By entirely removing Non-Maximum Suppression (NMS) post-processing, it ensures highly consistent latency and dramatically simplifies export pipelines for edge environments. +2. **Next-Generation Optimization:** Training stability is revolutionized by the **MuSGD Optimizer**, a hybrid of SGD and Muon (inspired by LLM methodologies like Moonshot AI's Kimi K2). This guarantees faster convergence. Furthermore, YOLO26 utilizes **ProgLoss + STAL** to drastically improve small-object recognition, a crucial feature for applications involving [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and robotics. 3. **Unmatched Hardware Efficiency:** By removing Distribution Focal Loss (DFL), YOLO26 drastically lowers memory requirements. It boasts up to **43% faster CPU inference**, making it the definitive choice for devices lacking dedicated [GPU](https://www.ultralytics.com/glossary/gpu-graphics-processing-unit) acceleration. -4. **Extreme Versatility:** Unlike PP-YOLOE+ which focuses strictly on detection, YOLO26 offers unified support across numerous tasks. It incorporates a specialized semantic segmentation loss for [instance segmentation](https://docs.ultralytics.com/tasks/segment/), Residual Log-Likelihood Estimation (RLE) for accurate [pose estimation](https://docs.ultralytics.com/tasks/pose/), and advanced angle loss mechanisms for [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). +4. **Extreme Versatility:** Unlike PP-YOLOE+ which focuses strictly on detection, YOLO26 offers unified support across numerous tasks. It incorporates a specialized semantic segmentation loss for [instance segmentation](https://docs.ultralytics.com/tasks/segment), Residual Log-Likelihood Estimation (RLE) for accurate [pose estimation](https://docs.ultralytics.com/tasks/pose), and advanced angle loss mechanisms for [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb). [Learn more about YOLO26](https://platform.ultralytics.com/ultralytics/yolo26){ .md-button } @@ -119,7 +119,7 @@ predictions = model("https://ultralytics.com/images/bus.jpg") model.export(format="onnx") ``` -For users evaluating other robust architectures within the Ultralytics ecosystem, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a highly reliable choice for legacy deployments, while the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) provides excellent capabilities for those seeking attention-based solutions. +For users evaluating other robust architectures within the Ultralytics ecosystem, [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) remains a highly reliable choice for legacy deployments, while the transformer-based [RT-DETR](https://docs.ultralytics.com/models/rtdetr) provides excellent capabilities for those seeking attention-based solutions. ## Summary diff --git a/docs/en/compare/yolox-vs-rtdetr.md b/docs/en/compare/yolox-vs-rtdetr.md index 1dfe61e9814..6b7bb5d708b 100644 --- a/docs/en/compare/yolox-vs-rtdetr.md +++ b/docs/en/compare/yolox-vs-rtdetr.md @@ -8,7 +8,7 @@ keywords: YOLOX, RTDETRv2, object detection, YOLOX vs RTDETRv2, performance comp Choosing the optimal architecture for [computer vision applications](https://www.ultralytics.com/glossary/computer-vision-cv) requires a careful balance of accuracy, inference speed, and deployment feasibility. In this comprehensive technical analysis, we explore the fundamental differences between **YOLOX**, a highly successful anchor-free CNN architecture, and **RTDETRv2**, a state-of-the-art real-time detection transformer. -While both models have made significant contributions to the field of [object detection](https://docs.ultralytics.com/tasks/detect/), developers building production-ready applications often find that modern alternatives like [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) provide superior training efficiency, lower memory requirements, and a more robust deployment ecosystem. +While both models have made significant contributions to the field of [object detection](https://docs.ultralytics.com/tasks/detect), developers building production-ready applications often find that modern alternatives like [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) provide superior training efficiency, lower memory requirements, and a more robust deployment ecosystem. @@ -47,13 +47,13 @@ Building upon the foundation of its predecessor, RTDETRv2 leverages the power of RTDETRv2 fundamentally reimagines the detection pipeline by utilizing a transformer-based architecture that natively bypasses Non-Maximum Suppression (NMS). This is achieved through a hybrid encoder and IoU-aware query selection, which improves the initialization of object queries. The model effectively handles multi-scale features, allowing it to capture intricate details in complex environments, such as [traffic video detection at nighttime](https://www.ultralytics.com/blog/traffic-video-detection-at-nighttime-a-look-at-why-accuracy-is-key). -However, transformers are inherently resource-intensive. Training RTDETRv2 typically demands significantly more GPU memory and compute cycles than CNN-based alternatives, which can be a hurdle for teams operating within strict budget constraints or those requiring frequent [model tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/). +However, transformers are inherently resource-intensive. Training RTDETRv2 typically demands significantly more GPU memory and compute cycles than CNN-based alternatives, which can be a hurdle for teams operating within strict budget constraints or those requiring frequent [model tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning). -[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr/){ .md-button } +[Learn more about RTDETR](https://docs.ultralytics.com/models/rtdetr){ .md-button } ## Performance Comparison Table -To objectively evaluate these architectures, we examine their performance on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco/). The table below illustrates the trade-offs between accuracy ([mAP](https://www.ultralytics.com/glossary/mean-average-precision-map)), parameter count, and computational complexity. +To objectively evaluate these architectures, we examine their performance on the [COCO dataset](https://docs.ultralytics.com/datasets/detect/coco). The table below illustrates the trade-offs between accuracy ([mAP](https://www.ultralytics.com/glossary/mean-average-precision-map)), parameter count, and computational complexity. | Model | size
(pixels) | mAPval
50-95 | Speed
CPU ONNX
(ms) | Speed
T4 TensorRT10
(ms) | params
(M) | FLOPs
(B) | | ---------- | --------------------------- | -------------------------- | ------------------------------------ | ----------------------------------------- | ------------------------ | ----------------------- | @@ -93,11 +93,11 @@ RT-DETR is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage: YOLO26 @@ -109,7 +109,7 @@ Taking inspiration from transformer models while retaining the efficiency of CNN ### 2. Up to 43% Faster CPU Inference -Unlike transformer architectures like RTDETRv2 which heavily rely on high-end GPUs, YOLO26 is specifically optimized for [edge computing environments](https://www.ultralytics.com/glossary/edge-computing). Through the removal of Distribution Focal Loss (DFL), YOLO26 streamlines model export and achieves up to 43% faster CPU inference, making it the ideal choice for integration into hardware like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or standard mobile devices. +Unlike transformer architectures like RTDETRv2 which heavily rely on high-end GPUs, YOLO26 is specifically optimized for [edge computing environments](https://www.ultralytics.com/glossary/edge-computing). Through the removal of Distribution Focal Loss (DFL), YOLO26 streamlines model export and achieves up to 43% faster CPU inference, making it the ideal choice for integration into hardware like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or standard mobile devices. ### 3. Training Efficiency with MuSGD @@ -117,7 +117,7 @@ Training transformer models often leads to excessive [CUDA memory consumption](h ### 4. Unmatched Ecosystem and Versatility -The [Ultralytics ecosystem](https://docs.ultralytics.com/) provides an intuitive, streamlined developer experience. With extensive documentation, active community support, and the cloud-powered [Ultralytics Platform](https://platform.ultralytics.com), managing the complete AI lifecycle has never been easier. Furthermore, YOLO26 is highly versatile. While RTDETRv2 focuses on object detection, YOLO26 seamlessly supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), [image classification](https://docs.ultralytics.com/tasks/classify/), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) tasks natively. Enhanced by the new **ProgLoss + STAL** loss functions, YOLO26 also excels at small-object recognition, a critical feature for [aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision) and [industrial defect detection](https://www.ultralytics.com/blog/how-vision-ai-enhances-defect-detection-on-production-lines). +The [Ultralytics ecosystem](https://docs.ultralytics.com/) provides an intuitive, streamlined developer experience. With extensive documentation, active community support, and the cloud-powered [Ultralytics Platform](https://platform.ultralytics.com), managing the complete AI lifecycle has never been easier. Furthermore, YOLO26 is highly versatile. While RTDETRv2 focuses on object detection, YOLO26 seamlessly supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), [image classification](https://docs.ultralytics.com/tasks/classify), and [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) tasks natively. Enhanced by the new **ProgLoss + STAL** loss functions, YOLO26 also excels at small-object recognition, a critical feature for [aerial imagery](https://www.ultralytics.com/blog/12-aerial-imagery-use-cases-powered-by-computer-vision) and [industrial defect detection](https://www.ultralytics.com/blog/how-vision-ai-enhances-defect-detection-on-production-lines). !!! tip "Other Supported Models" diff --git a/docs/en/compare/yolox-vs-yolo11.md b/docs/en/compare/yolox-vs-yolo11.md index 1ea7f5c5cf2..adaca1f5dc4 100644 --- a/docs/en/compare/yolox-vs-yolo11.md +++ b/docs/en/compare/yolox-vs-yolo11.md @@ -39,7 +39,7 @@ Released on September 27, 2024, by Glenn Jocher and Jing Qiu at [Ultralytics](ht ### The Ultralytics Advantage -YOLO11 is not just an object detector; it is a unified framework supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection. It boasts a highly efficient architecture that prioritizes a seamless balance between speed, parameter count, and accuracy. +YOLO11 is not just an object detector; it is a unified framework supporting [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding box (OBB)](https://docs.ultralytics.com/tasks/obb) detection. It boasts a highly efficient architecture that prioritizes a seamless balance between speed, parameter count, and accuracy. Furthermore, YOLO11 is fully integrated into the [Ultralytics Platform](https://platform.ultralytics.com/ultralytics/yolo11), which provides a streamlined ecosystem for data annotation, model training, and deployment. @@ -64,7 +64,7 @@ When comparing these models, the balance of performance becomes clear. YOLO11 ac | YOLO11l | 640 | 53.4 | 238.6 | 6.2 | 25.3 | 86.9 | | YOLO11x | 640 | **54.7** | 462.8 | 11.3 | 56.9 | 194.9 | -As demonstrated, YOLO11 models consistently outperform YOLOX in accuracy while maintaining a leaner parameter footprint. For instance, YOLO11m achieves a **51.5 mAP** with only **20.1M parameters**, whereas YOLOXx achieves a similar 51.1 mAP but requires a massive **99.1M parameters**. This memory efficiency during training and inference makes YOLO11 highly suitable for deployment on edge AI devices, avoiding the heavy CUDA memory requirements typical of older or transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). +As demonstrated, YOLO11 models consistently outperform YOLOX in accuracy while maintaining a leaner parameter footprint. For instance, YOLO11m achieves a **51.5 mAP** with only **20.1M parameters**, whereas YOLOXx achieves a similar 51.1 mAP but requires a massive **99.1M parameters**. This memory efficiency during training and inference makes YOLO11 highly suitable for deployment on edge AI devices, avoiding the heavy CUDA memory requirements typical of older or transformer-based models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). !!! tip "Efficient Training" @@ -76,7 +76,7 @@ One of the most striking differences between the two frameworks is the developer YOLOX often requires cloning repositories, setting up complex environments, and running verbose command-line arguments to train and export models to formats like [ONNX](https://onnx.ai/) or [TensorRT](https://developer.nvidia.com/tensorrt). -In stark contrast, **Ultralytics YOLO11** offers an incredibly simple Python API and CLI. The Ultralytics library handles [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation/), [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning/), and exporting automatically. +In stark contrast, **Ultralytics YOLO11** offers an incredibly simple Python API and CLI. The Ultralytics library handles [data augmentation](https://docs.ultralytics.com/guides/yolo-data-augmentation), [hyperparameter tuning](https://docs.ultralytics.com/guides/hyperparameter-tuning), and exporting automatically. ```python from ultralytics import YOLO @@ -91,7 +91,7 @@ results = model.train(data="coco8.yaml", epochs=100, imgsz=640) model.export(format="engine") ``` -This well-maintained ecosystem is backed by extensive [documentation](https://docs.ultralytics.com/) and seamless integration with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) for [experiment tracking](https://www.ultralytics.com/glossary/experiment-tracking). +This well-maintained ecosystem is backed by extensive [documentation](https://docs.ultralytics.com/) and seamless integration with tools like [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) for [experiment tracking](https://www.ultralytics.com/glossary/experiment-tracking). ## Ideal Use Cases @@ -104,9 +104,9 @@ Choosing between these models often depends on the specifics of the deployment e ### When to use YOLO11 -- **Production Deployments:** For commercial applications in [smart retail](https://www.ultralytics.com/solutions/ai-in-retail) or [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/), where robust, maintained code and high accuracy are non-negotiable. +- **Production Deployments:** For commercial applications in [smart retail](https://www.ultralytics.com/solutions/ai-in-retail) or [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system), where robust, maintained code and high accuracy are non-negotiable. - **Multi-Task Pipelines:** When a project requires tracking objects, estimating human poses, and segmenting instances using a single, unified framework. -- **Resource-Constrained Edge Devices:** Because of its low parameter count and high throughput, YOLO11 is ideal for deployment on [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or mobile edge nodes via [CoreML](https://docs.ultralytics.com/integrations/coreml/) and [NCNN](https://docs.ultralytics.com/integrations/ncnn/). +- **Resource-Constrained Edge Devices:** Because of its low parameter count and high throughput, YOLO11 is ideal for deployment on [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or mobile edge nodes via [CoreML](https://docs.ultralytics.com/integrations/coreml) and [NCNN](https://docs.ultralytics.com/integrations/ncnn). ## Looking Ahead: The YOLO26 Advantage @@ -114,9 +114,9 @@ While YOLO11 represents a massive leap over YOLOX, the field of computer vision Released in January 2026, YOLO26 takes the architectural brilliance of YOLO11 and introduces several groundbreaking features: -- **End-to-End NMS-Free Design:** YOLO26 eliminates Non-Maximum Suppression (NMS) post-processing, natively streaming inference for faster, simpler deployment pipelines (a concept first explored in [YOLOv10](https://docs.ultralytics.com/models/yolov10/)). +- **End-to-End NMS-Free Design:** YOLO26 eliminates Non-Maximum Suppression (NMS) post-processing, natively streaming inference for faster, simpler deployment pipelines (a concept first explored in [YOLOv10](https://docs.ultralytics.com/models/yolov10)). - **Up to 43% Faster CPU Inference:** Through the removal of Distribution Focal Loss (DFL), YOLO26 is vastly more efficient on CPUs and low-power edge devices. - **MuSGD Optimizer:** Inspired by LLM training innovations from Moonshot AI, the MuSGD optimizer ensures highly stable training runs and rapid convergence. -- **Advanced Loss Functions:** Utilizing ProgLoss + STAL, YOLO26 achieves notable improvements in small-object recognition, which is critical for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and autonomous robotics. +- **Advanced Loss Functions:** Utilizing ProgLoss + STAL, YOLO26 achieves notable improvements in small-object recognition, which is critical for [drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and autonomous robotics. -For the vast majority of modern computer vision tasks, upgrading your pipeline to leverage [YOLO26](https://docs.ultralytics.com/models/yolo26/) will provide the absolute best balance of speed, accuracy, and deployment simplicity. +For the vast majority of modern computer vision tasks, upgrading your pipeline to leverage [YOLO26](https://docs.ultralytics.com/models/yolo26) will provide the absolute best balance of speed, accuracy, and deployment simplicity. diff --git a/docs/en/compare/yolox-vs-yolo26.md b/docs/en/compare/yolox-vs-yolo26.md index af531eb92f8..e61cb30659e 100644 --- a/docs/en/compare/yolox-vs-yolo26.md +++ b/docs/en/compare/yolox-vs-yolo26.md @@ -26,7 +26,7 @@ Arxiv: [2107.08430](https://arxiv.org/abs/2107.08430) GitHub: [Megvii-BaseDetection/YOLOX](https://github.com/Megvii-BaseDetection/YOLOX) Docs: [YOLOX ReadTheDocs](https://yolox.readthedocs.io/en/latest/) -Introduced in mid-2021, YOLOX represented a major shift by adopting an anchor-free design coupled with a decoupled head and the advanced label assignment strategy known as SimOTA. By stepping away from the traditional anchor box mechanisms that dominated previous architectures, YOLOX successfully bridged the gap between academic research and industrial application, offering an elegant yet highly effective framework for [object detection](https://docs.ultralytics.com/tasks/detect/). +Introduced in mid-2021, YOLOX represented a major shift by adopting an anchor-free design coupled with a decoupled head and the advanced label assignment strategy known as SimOTA. By stepping away from the traditional anchor box mechanisms that dominated previous architectures, YOLOX successfully bridged the gap between academic research and industrial application, offering an elegant yet highly effective framework for [object detection](https://docs.ultralytics.com/tasks/detect). [Learn more about YOLOX](https://github.com/Megvii-BaseDetection/YOLOX){ .md-button } @@ -57,8 +57,8 @@ YOLO26 takes architectural efficiency to the next level. The removal of NMS not Key YOLO26 innovations include: - **MuSGD Optimizer:** Inspired by Large Language Model (LLM) training techniques, this hybrid of SGD and Muon ensures exceptionally stable training runs and faster convergence. -- **Up to 43% Faster CPU Inference:** By eliminating DFL and streamlining the network architecture, YOLO26 is heavily optimized for resource-constrained edge devices, from simple IoT sensors to [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) boards. -- **ProgLoss + STAL:** These advanced loss functions deliver notable improvements in small-object recognition, which is critical for analyzing [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and performing precise quality control in [manufacturing automation](https://www.ultralytics.com/blog/manufacturing-automation). +- **Up to 43% Faster CPU Inference:** By eliminating DFL and streamlining the network architecture, YOLO26 is heavily optimized for resource-constrained edge devices, from simple IoT sensors to [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) boards. +- **ProgLoss + STAL:** These advanced loss functions deliver notable improvements in small-object recognition, which is critical for analyzing [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and performing precise quality control in [manufacturing automation](https://www.ultralytics.com/blog/manufacturing-automation). !!! tip "Edge-First Optimization" @@ -91,7 +91,7 @@ One of the most profound differences between these architectures lies in their u While YOLOX remains a foundational repository for researchers studying gradient flow and anchor-free mechanics, its setup can be complex, often requiring manual configuration of dependencies and operators. Conversely, the **[Ultralytics ecosystem](https://docs.ultralytics.com/)** defines the industry standard for ease of use. -By utilizing the unified Python API, developers can initialize, train, and deploy YOLO26 models with unparalleled simplicity. The system inherently handles dataset downloading, hyperparameter tuning, and seamless export to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), and OpenVINO. +By utilizing the unified Python API, developers can initialize, train, and deploy YOLO26 models with unparalleled simplicity. The system inherently handles dataset downloading, hyperparameter tuning, and seamless export to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx), [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), and OpenVINO. ```python from ultralytics import YOLO @@ -121,7 +121,7 @@ YOLOX remains a viable candidate for specific academic benchmarks and legacy sys ### Where YOLO26 Excels -YOLO26 is fundamentally designed for modern industrial applications. Because it natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), it is far more versatile than standard detection engines. +YOLO26 is fundamentally designed for modern industrial applications. Because it natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb), it is far more versatile than standard detection engines. - **Smart Retail and Inventory:** Utilizing the NMS-free design guarantees that automated checkout systems process video feeds with ultra-low latency, recognizing products without the bottleneck of post-processing loops. - **Drone and Aerial Analytics:** The specialized angle loss for OBB and the integration of ProgLoss + STAL make YOLO26 unmatched at detecting rotated objects and tiny artifacts in vast satellite images. @@ -145,7 +145,7 @@ YOLO26 is recommended for: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Exploring Other Ultralytics Models diff --git a/docs/en/compare/yolox-vs-yolov10.md b/docs/en/compare/yolox-vs-yolov10.md index e4641c0e4b0..c8a32425a2d 100644 --- a/docs/en/compare/yolox-vs-yolov10.md +++ b/docs/en/compare/yolox-vs-yolov10.md @@ -35,9 +35,9 @@ Organization: [Tsinghua University](https://www.tsinghua.edu.cn/en/) Date: 2024-05-23 Arxiv: [https://arxiv.org/abs/2405.14458](https://arxiv.org/abs/2405.14458) GitHub: [https://github.com/THU-MIG/yolov10](https://github.com/THU-MIG/yolov10) -Docs: [https://docs.ultralytics.com/models/yolov10/](https://docs.ultralytics.com/models/yolov10/) +Docs: [https://docs.ultralytics.com/models/yolov10/](https://docs.ultralytics.com/models/yolov10) -[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10/){ .md-button } +[Learn more about YOLOv10](https://docs.ultralytics.com/models/yolov10){ .md-button } ## Architectural Innovations @@ -53,7 +53,7 @@ While YOLOX simplified the detection head, it still relied on NMS to filter out !!! note "The Impact of Removing NMS" - Non-Maximum Suppression is often a complex operation to accelerate on Neural Processing Units (NPUs). By removing it, YOLOv10 allows the entire model graph to execute seamlessly on specialized hardware, drastically improving compatibility with optimization frameworks like [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) and TensorRT. + Non-Maximum Suppression is often a complex operation to accelerate on Neural Processing Units (NPUs). By removing it, YOLOv10 allows the entire model graph to execute seamlessly on specialized hardware, drastically improving compatibility with optimization frameworks like [OpenVINO](https://docs.ultralytics.com/integrations/openvino) and TensorRT. ## Performance Metrics and Comparison @@ -77,13 +77,13 @@ When evaluating models for production, balancing accuracy with computational ove ### Analyzing the Data -The metrics clearly demonstrate YOLOv10's generational leap. For instance, YOLOv10-S achieves a [mean Average Precision](https://docs.ultralytics.com/guides/yolo-performance-metrics/) of 46.7% compared to YOLOX-m's 46.9%, but does so using less than a third of the parameters (7.2M vs 25.3M) and significantly fewer FLOPs. Furthermore, the top-tier YOLOv10-X model pushes the mAP to 54.4%, making it highly competitive for demanding accuracy tasks while remaining faster than the older YOLOX-x architecture. +The metrics clearly demonstrate YOLOv10's generational leap. For instance, YOLOv10-S achieves a [mean Average Precision](https://docs.ultralytics.com/guides/yolo-performance-metrics) of 46.7% compared to YOLOX-m's 46.9%, but does so using less than a third of the parameters (7.2M vs 25.3M) and significantly fewer FLOPs. Furthermore, the top-tier YOLOv10-X model pushes the mAP to 54.4%, making it highly competitive for demanding accuracy tasks while remaining faster than the older YOLOX-x architecture. ## The Ultralytics Ecosystem Advantage While YOLOX remains a robust open-source research implementation, adopting YOLOv10 provides immediate access to the well-maintained ecosystem provided by Ultralytics. Choosing an Ultralytics-supported model ensures a streamlined user experience characterized by a simple API and extensive documentation. -Developers benefit heavily from the framework's memory requirements; training Ultralytics models typically consumes far less CUDA memory than heavy transformer-based alternatives like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). This efficient training footprint allows for larger batch sizes on consumer-grade hardware, accelerating the time from data collection to model deployment. Furthermore, the framework offers unmatched versatility, allowing users to switch seamlessly between [object detection](https://docs.ultralytics.com/tasks/detect/), [instance segmentation](https://docs.ultralytics.com/tasks/segment/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) with minimal code changes. +Developers benefit heavily from the framework's memory requirements; training Ultralytics models typically consumes far less CUDA memory than heavy transformer-based alternatives like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). This efficient training footprint allows for larger batch sizes on consumer-grade hardware, accelerating the time from data collection to model deployment. Furthermore, the framework offers unmatched versatility, allowing users to switch seamlessly between [object detection](https://docs.ultralytics.com/tasks/detect), [instance segmentation](https://docs.ultralytics.com/tasks/segment), and [pose estimation](https://docs.ultralytics.com/tasks/pose) with minimal code changes. ### Training and Inference Example @@ -105,7 +105,7 @@ predictions = model.predict("https://ultralytics.com/images/bus.jpg") model.export(format="engine", half=True) ``` -By leveraging built-in export routines, converting models to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) or [ONNX](https://docs.ultralytics.com/integrations/onnx/) requires just a single line of code, entirely bypassing complex compilation hurdles. +By leveraging built-in export routines, converting models to formats like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) or [ONNX](https://docs.ultralytics.com/integrations/onnx) requires just a single line of code, entirely bypassing complex compilation hurdles. ## Ideal Use Cases and Deployment Scenarios diff --git a/docs/en/compare/yolox-vs-yolov5.md b/docs/en/compare/yolox-vs-yolov5.md index 3f9de497b50..19a8eba4381 100644 --- a/docs/en/compare/yolox-vs-yolov5.md +++ b/docs/en/compare/yolox-vs-yolov5.md @@ -76,19 +76,19 @@ When transitioning from research to production, the ecosystem surrounding a mode ### Streamlined User Experience -YOLOv5 is universally praised for its "zero-to-hero" developer experience. The [Ultralytics Python API](https://docs.ultralytics.com/usage/python/) and CLI allow you to load, train, and deploy models with single lines of code. In contrast, running YOLOX from the [Megvii GitHub repository](https://github.com/Megvii-BaseDetection/YOLOX) requires more manual configuration of environment variables, complex Python path setups, and a steeper learning curve typical of academic research codebases. +YOLOv5 is universally praised for its "zero-to-hero" developer experience. The [Ultralytics Python API](https://docs.ultralytics.com/usage/python) and CLI allow you to load, train, and deploy models with single lines of code. In contrast, running YOLOX from the [Megvii GitHub repository](https://github.com/Megvii-BaseDetection/YOLOX) requires more manual configuration of environment variables, complex Python path setups, and a steeper learning curve typical of academic research codebases. ### Training Efficiency and Memory Requirements -Ultralytics models are meticulously engineered to minimize memory usage during training. YOLOv5 requires significantly less CUDA memory compared to heavily parameterized transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) or unoptimized research models. This allows developers to train larger batch sizes on consumer-grade hardware, accelerating the iterative development cycle. +Ultralytics models are meticulously engineered to minimize memory usage during training. YOLOv5 requires significantly less CUDA memory compared to heavily parameterized transformer models like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) or unoptimized research models. This allows developers to train larger batch sizes on consumer-grade hardware, accelerating the iterative development cycle. ### Versatility Across Tasks -While YOLOX is strictly an object detection framework, the Ultralytics ecosystem has evolved YOLOv5 to support multiple vision tasks. Out of the box, you can perform [Image Classification](https://docs.ultralytics.com/tasks/classify/), [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), and object detection using the exact same API syntax. +While YOLOX is strictly an object detection framework, the Ultralytics ecosystem has evolved YOLOv5 to support multiple vision tasks. Out of the box, you can perform [Image Classification](https://docs.ultralytics.com/tasks/classify), [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), and object detection using the exact same API syntax. !!! tip "Continuous Innovation" - If you require even more advanced tasks like [Pose Estimation](https://docs.ultralytics.com/tasks/pose/) or [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb/) detection, we highly recommend upgrading to the latest [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) architecture, which supports all these natively with state-of-the-art accuracy. + If you require even more advanced tasks like [Pose Estimation](https://docs.ultralytics.com/tasks/pose) or [Oriented Bounding Box (OBB)](https://docs.ultralytics.com/tasks/obb) detection, we highly recommend upgrading to the latest [Ultralytics YOLO26](https://platform.ultralytics.com/ultralytics/yolo26) architecture, which supports all these natively with state-of-the-art accuracy. ## Code Comparison @@ -117,7 +117,7 @@ _(Requires manual repository cloning, setup.py installation, and complex CLI arg python tools/train.py -f exps/default/yolox_s.py -d 1 -b 64 --fp16 -o ``` -The Ultralytics approach removes friction, allowing you to focus on your dataset and application logic rather than debugging configuration files. Furthermore, tracking your experiments is seamless with built-in integrations for [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases/) and [Comet ML](https://docs.ultralytics.com/integrations/comet/). +The Ultralytics approach removes friction, allowing you to focus on your dataset and application logic rather than debugging configuration files. Furthermore, tracking your experiments is seamless with built-in integrations for [Weights & Biases](https://docs.ultralytics.com/integrations/weights-biases) and [Comet ML](https://docs.ultralytics.com/integrations/comet). ## Ideal Use Cases and Real-World Applications @@ -133,7 +133,7 @@ YOLOv5 is the undisputed champion of practical deployment. - **High-Speed Manufacturing:** For assembly line [defect detection](https://www.ultralytics.com/blog/how-vision-ai-enhances-defect-detection-on-production-lines), YOLOv5's minimal inference latency on edge GPUs ensures products are inspected without slowing down the belt. - **Drone and Aerial Imagery:** Its efficient memory footprint allows it to run on lightweight companion computers on drones for tasks like [agriculture monitoring](https://www.ultralytics.com/solutions/ai-in-agriculture) and wildlife tracking. -- **Smart Retail:** From [automated checkout](https://www.ultralytics.com/solutions/ai-in-retail) to inventory management, YOLOv5 easily exports to [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) and [ONNX](https://docs.ultralytics.com/integrations/onnx/) for mass deployment across thousands of store cameras. +- **Smart Retail:** From [automated checkout](https://www.ultralytics.com/solutions/ai-in-retail) to inventory management, YOLOv5 easily exports to [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) and [ONNX](https://docs.ultralytics.com/integrations/onnx) for mass deployment across thousands of store cameras. ## Looking Forward: The YOLO26 Advantage diff --git a/docs/en/compare/yolox-vs-yolov6.md b/docs/en/compare/yolox-vs-yolov6.md index 12961ae3df5..34dd45f4027 100644 --- a/docs/en/compare/yolox-vs-yolov6.md +++ b/docs/en/compare/yolox-vs-yolov6.md @@ -39,7 +39,7 @@ Furthermore, YOLOX introduced a decoupled head architecture. By separating the c ## YOLOv6-3.0: The Industrial Heavyweight -Developed by the Vision AI Department at [Meituan](https://www.meituan.com/), YOLOv6-3.0 is unapologetically engineered for maximum industrial throughput, particularly on NVIDIA GPUs using hardware accelerators like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/). +Developed by the Vision AI Department at [Meituan](https://www.meituan.com/), YOLOv6-3.0 is unapologetically engineered for maximum industrial throughput, particularly on NVIDIA GPUs using hardware accelerators like [TensorRT](https://docs.ultralytics.com/integrations/tensorrt). - **Authors:** Chuyi Li, Lulu Li, Yifei Geng, et al. - **Organization:** Meituan @@ -53,7 +53,7 @@ YOLOv6-3.0 focuses on maximizing [GPU](https://www.ultralytics.com/glossary/gpu- The backbone is constructed using the hardware-friendly EfficientRep architecture, deliberately designed to minimize memory access costs and maximize computational density on modern accelerators. This makes YOLOv6 an exceptionally strong candidate for server-side video analytics. -[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6/){ .md-button } +[Learn more about YOLOv6](https://docs.ultralytics.com/models/yolov6){ .md-button } ## Performance Comparison @@ -97,19 +97,19 @@ YOLOv6 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage While both Megvii and Meituan provide powerful research repositories, deploying these models in production often requires significant engineering overhead. The integrated [Ultralytics ecosystem](https://docs.ultralytics.com/) eliminates these hurdles by offering a unified, extensively documented API. -By leveraging the Ultralytics package, developers gain access to an unparalleled user experience. This includes built-in [auto-augmentation](https://docs.ultralytics.com/reference/data/augment/), highly efficient memory management during training (drastically lowering VRAM requirements compared to transformer models like [RTDETR](https://docs.ultralytics.com/models/rtdetr/)), and seamless export pipelines to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino/). +By leveraging the Ultralytics package, developers gain access to an unparalleled user experience. This includes built-in [auto-augmentation](https://docs.ultralytics.com/reference/data/augment), highly efficient memory management during training (drastically lowering VRAM requirements compared to transformer models like [RTDETR](https://docs.ultralytics.com/models/rtdetr)), and seamless export pipelines to formats like [ONNX](https://docs.ultralytics.com/integrations/onnx) and [OpenVINO](https://docs.ultralytics.com/integrations/openvino). -Unlike specialized models, Ultralytics architectures are inherently versatile, supporting [Object Detection](https://docs.ultralytics.com/tasks/detect/), [Instance Segmentation](https://docs.ultralytics.com/tasks/segment/), [Pose Estimation](https://docs.ultralytics.com/tasks/pose/), Image Classification, and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) out of the box. +Unlike specialized models, Ultralytics architectures are inherently versatile, supporting [Object Detection](https://docs.ultralytics.com/tasks/detect), [Instance Segmentation](https://docs.ultralytics.com/tasks/segment), [Pose Estimation](https://docs.ultralytics.com/tasks/pose), Image Classification, and [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb) out of the box. ### Enter YOLO26: The Ultimate Edge Solution diff --git a/docs/en/compare/yolox-vs-yolov7.md b/docs/en/compare/yolox-vs-yolov7.md index e3b788d2136..c00d5d0c2b9 100644 --- a/docs/en/compare/yolox-vs-yolov7.md +++ b/docs/en/compare/yolox-vs-yolov7.md @@ -6,7 +6,7 @@ keywords: YOLOX, YOLOv7, object detection, computer vision, model comparison, an # YOLOX vs YOLOv7: A Comprehensive Technical Comparison -The evolution of real-time [object detection](https://docs.ultralytics.com/tasks/detect/) has been driven by continuous architectural breakthroughs. Two significant milestones in this journey are **YOLOX** and **YOLOv7**. Released within a year of each other, both models introduced novel approaches to the standard object detection paradigm, significantly improving the trade-off between speed and [accuracy](https://www.ultralytics.com/glossary/accuracy). +The evolution of real-time [object detection](https://docs.ultralytics.com/tasks/detect) has been driven by continuous architectural breakthroughs. Two significant milestones in this journey are **YOLOX** and **YOLOv7**. Released within a year of each other, both models introduced novel approaches to the standard object detection paradigm, significantly improving the trade-off between speed and [accuracy](https://www.ultralytics.com/glossary/accuracy). This page provides an in-depth technical analysis of YOLOX and YOLOv7, comparing their architectures, performance metrics, and ideal use cases to help developers choose the right tool for their computer vision deployments. @@ -49,13 +49,13 @@ Released in July 2022 by researchers at the Institute of Information Science, Ac - **Date:** 2022-07-06 - **Research Paper:** [arXiv:2207.02696](https://arxiv.org/abs/2207.02696) - **Source Code:** [WongKinYiu YOLOv7 GitHub](https://github.com/WongKinYiu/yolov7) -- **Documentation:** [Ultralytics YOLOv7 Docs](https://docs.ultralytics.com/models/yolov7/) +- **Documentation:** [Ultralytics YOLOv7 Docs](https://docs.ultralytics.com/models/yolov7) ### Architectural Innovations YOLOv7's architecture is built around the Extended Efficient Layer Aggregation Network (E-ELAN), which allows the model to learn more diverse features continuously without degrading the gradient path. Furthermore, YOLOv7 utilized model re-parameterization techniques, enabling complex multi-branch training networks to be simplified into faster, single-path networks during inference. -[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7/){ .md-button } +[Learn more about YOLOv7](https://docs.ultralytics.com/models/yolov7){ .md-button } ## Performance Comparison @@ -76,8 +76,8 @@ When evaluating these models for real-world applications, understanding their pe ### Analysis - **Accuracy:** YOLOv7 generally achieves a higher [mAP](https://www.ultralytics.com/glossary/mean-average-precision-map) compared to the equivalent YOLOX models. For instance, YOLOv7x achieves 53.1 mAP compared to YOLOXx's 51.1. -- **Speed:** While both models are highly optimized for GPU execution using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/), YOLOv7's E-ELAN architecture provides slightly better throughput for high-end applications, though YOLOX maintains excellent latency on smaller edge devices. -- **Versatility:** YOLOv7 expanded its repertoire beyond bounding boxes by natively providing weights for [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [pose estimation](https://docs.ultralytics.com/tasks/pose/), making it more versatile than the base YOLOX repository. +- **Speed:** While both models are highly optimized for GPU execution using [TensorRT](https://docs.ultralytics.com/integrations/tensorrt), YOLOv7's E-ELAN architecture provides slightly better throughput for high-end applications, though YOLOX maintains excellent latency on smaller edge devices. +- **Versatility:** YOLOv7 expanded its repertoire beyond bounding boxes by natively providing weights for [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [pose estimation](https://docs.ultralytics.com/tasks/pose), making it more versatile than the base YOLOX repository. ## Real-World Applications @@ -85,7 +85,7 @@ Choosing between these models often comes down to your specific deployment envir ### Edge Computing and IoT -For constrained edge devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/) or older mobile processors, **YOLOX-Nano** and **YOLOX-Tiny** are highly attractive. Their minimal parameter count and anchor-free nature make them easier to deploy in low-power environments for tasks like basic motion tracking or smart doorbell applications. +For constrained edge devices like [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi) or older mobile processors, **YOLOX-Nano** and **YOLOX-Tiny** are highly attractive. Their minimal parameter count and anchor-free nature make them easier to deploy in low-power environments for tasks like basic motion tracking or smart doorbell applications. ### High-Fidelity Video Analytics @@ -113,11 +113,11 @@ YOLOv7 is recommended for: ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## The Ultralytics Advantage @@ -125,11 +125,11 @@ While both YOLOX and YOLOv7 are powerful research implementations, moving from a Ultralytics models provide a **unified Python API**, treating model training, validation, and deployment as streamlined, standardized tasks. You avoid the headache of managing complex third-party dependencies or custom C++ operators common in older architectures. -Furthermore, Ultralytics YOLO models require significantly **less CUDA memory** during training compared to transformer-based detectors like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/). This allows practitioners to utilize larger [batch sizes](https://www.ultralytics.com/glossary/batch-size), stabilizing training and accelerating convergence on custom datasets. +Furthermore, Ultralytics YOLO models require significantly **less CUDA memory** during training compared to transformer-based detectors like [RT-DETR](https://docs.ultralytics.com/models/rtdetr). This allows practitioners to utilize larger [batch sizes](https://www.ultralytics.com/glossary/batch-size), stabilizing training and accelerating convergence on custom datasets. !!! tip "Supported Integrations" - Ultralytics natively supports exporting models to industry-standard formats like [ONNX](https://docs.ultralytics.com/integrations/onnx/), [OpenVINO](https://docs.ultralytics.com/integrations/openvino/), and [CoreML](https://docs.ultralytics.com/integrations/coreml/) with a simple boolean flag, vastly simplifying the [model deployment process](https://docs.ultralytics.com/guides/model-deployment-options/). + Ultralytics natively supports exporting models to industry-standard formats like [ONNX](https://docs.ultralytics.com/integrations/onnx), [OpenVINO](https://docs.ultralytics.com/integrations/openvino), and [CoreML](https://docs.ultralytics.com/integrations/coreml) with a simple boolean flag, vastly simplifying the [model deployment process](https://docs.ultralytics.com/guides/model-deployment-options). ## Code Example: Training with Ultralytics @@ -159,7 +159,7 @@ While YOLOv7 and YOLOX represent important historical steps, the state-of-the-ar - **End-to-End NMS-Free Design:** YOLO26 natively eliminates [Non-Maximum Suppression (NMS)](https://www.ultralytics.com/glossary/non-maximum-suppression-nms) post-processing. This drastically reduces latency bottlenecks and guarantees deterministic execution times across varied hardware setups. - **Up to 43% Faster CPU Inference:** By removing Distribution Focal Loss (DFL) and optimizing network depth, YOLO26 is heavily tailored for edge devices lacking dedicated GPU hardware. - **MuSGD Optimizer:** Inspired by advanced LLM training techniques, the MuSGD optimizer (a hybrid of SGD and Muon) offers exceptional training stability and faster convergence. -- **Improved Small Object Detection:** The integration of the ProgLoss + STAL loss functions provides significant improvements in recognizing small, distant objects—critical for [drone mapping](https://docs.ultralytics.com/datasets/detect/visdrone/) and security surveillance. -- **Native Task Support:** YOLO26 comprehensively supports [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/), instance segmentation, and pose estimation natively within the same streamlined API. +- **Improved Small Object Detection:** The integration of the ProgLoss + STAL loss functions provides significant improvements in recognizing small, distant objects—critical for [drone mapping](https://docs.ultralytics.com/datasets/detect/visdrone) and security surveillance. +- **Native Task Support:** YOLO26 comprehensively supports [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb), instance segmentation, and pose estimation natively within the same streamlined API. -For any modern developer starting a new computer vision project today, evaluating [Ultralytics YOLO26 on the Platform](https://platform.ultralytics.com/ultralytics/yolo26) is the recommended path to achieving the absolute best balance of speed, accuracy, and deployment simplicity. For those upgrading from previous generations like [YOLO11](https://docs.ultralytics.com/models/yolo11/) or [YOLOv8](https://docs.ultralytics.com/models/yolov8/), the transition requires changing only the model string, instantly unlocking superior capabilities. +For any modern developer starting a new computer vision project today, evaluating [Ultralytics YOLO26 on the Platform](https://platform.ultralytics.com/ultralytics/yolo26) is the recommended path to achieving the absolute best balance of speed, accuracy, and deployment simplicity. For those upgrading from previous generations like [YOLO11](https://docs.ultralytics.com/models/yolo11) or [YOLOv8](https://docs.ultralytics.com/models/yolov8), the transition requires changing only the model string, instantly unlocking superior capabilities. diff --git a/docs/en/compare/yolox-vs-yolov8.md b/docs/en/compare/yolox-vs-yolov8.md index 6bcd817045f..4ae39eb22d8 100644 --- a/docs/en/compare/yolox-vs-yolov8.md +++ b/docs/en/compare/yolox-vs-yolov8.md @@ -31,7 +31,7 @@ Docs: [YOLOX Documentation](https://yolox.readthedocs.io/en/latest/) YOLOX integrates several key modifications that set it apart from its predecessors. The most notable is the decoupled head, which separates classification and bounding box regression tasks into distinct pathways. This architectural choice resolves the inherent conflict between spatial alignment needed for regression and translation invariance required for classification, leading to a faster convergence rate during training. -Furthermore, YOLOX employs the SimOTA label assignment strategy. This dynamic assignment method formulates the matching of ground truth objects to predictions as an optimal transport problem, effectively reducing training time while boosting [mean average precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics/). The model also utilizes strong data augmentation techniques, including MixUp and Mosaic, though it notably turns them off during the final epochs to stabilize the learned features. +Furthermore, YOLOX employs the SimOTA label assignment strategy. This dynamic assignment method formulates the matching of ground truth objects to predictions as an optimal transport problem, effectively reducing training time while boosting [mean average precision (mAP)](https://docs.ultralytics.com/guides/yolo-performance-metrics). The model also utilizes strong data augmentation techniques, including MixUp and Mosaic, though it notably turns them off during the final epochs to stabilize the learned features. [Learn more about YOLOX](https://github.com/Megvii-BaseDetection/YOLOX){ .md-button } @@ -44,7 +44,7 @@ Author: Glenn Jocher, Ayush Chaurasia, and Jing Qiu Organization: [Ultralytics](https://www.ultralytics.com) Date: 2023-01-10 GitHub: [ultralytics/ultralytics](https://github.com/ultralytics/ultralytics) -Docs: [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8/) +Docs: [YOLOv8 Documentation](https://docs.ultralytics.com/models/yolov8) ### Architectural Advancements @@ -52,9 +52,9 @@ YOLOv8 introduces a streamlined architecture that replaces the C3 module with th !!! tip "The Ultralytics Ecosystem" - One of the greatest strengths of YOLOv8 is its deep integration into the Ultralytics ecosystem. Whether you are using the unified Python API or the visual interface of the [Ultralytics Platform](https://platform.ultralytics.com/), the transition from training to deployment is seamless, supporting formats from [ONNX](https://docs.ultralytics.com/integrations/onnx/) to [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) natively. + One of the greatest strengths of YOLOv8 is its deep integration into the Ultralytics ecosystem. Whether you are using the unified Python API or the visual interface of the [Ultralytics Platform](https://platform.ultralytics.com/), the transition from training to deployment is seamless, supporting formats from [ONNX](https://docs.ultralytics.com/integrations/onnx) to [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) natively. -Beyond standard [object detection](https://docs.ultralytics.com/tasks/detect/), YOLOv8 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment/), [image classification](https://docs.ultralytics.com/tasks/classify/), [pose estimation](https://docs.ultralytics.com/tasks/pose/), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb/). This multi-task versatility makes it a highly attractive choice for complex production environments where multiple model types must be maintained. +Beyond standard [object detection](https://docs.ultralytics.com/tasks/detect), YOLOv8 natively supports [instance segmentation](https://docs.ultralytics.com/tasks/segment), [image classification](https://docs.ultralytics.com/tasks/classify), [pose estimation](https://docs.ultralytics.com/tasks/pose), and [oriented bounding boxes (OBB)](https://docs.ultralytics.com/tasks/obb). This multi-task versatility makes it a highly attractive choice for complex production environments where multiple model types must be maintained. [Learn more about YOLOv8](https://platform.ultralytics.com/ultralytics/yolov8){ .md-button } @@ -77,7 +77,7 @@ When comparing these models, developers must consider the trade-offs between pre | YOLOv8l | 640 | 52.9 | 375.2 | 9.06 | 43.7 | 165.2 | | YOLOv8x | 640 | **53.9** | 479.1 | 14.37 | 68.2 | 257.8 | -YOLOv8 consistently demonstrates superior mAP across comparable parameter sizes while maintaining excellent GPU speeds. Furthermore, the Ultralytics models are known for their lower memory requirements during training. This is a crucial advantage when scaling batch sizes on consumer hardware, particularly when contrasted with resource-heavy transformer architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr/) which consume significantly more CUDA memory. +YOLOv8 consistently demonstrates superior mAP across comparable parameter sizes while maintaining excellent GPU speeds. Furthermore, the Ultralytics models are known for their lower memory requirements during training. This is a crucial advantage when scaling batch sizes on consumer hardware, particularly when contrasted with resource-heavy transformer architectures like [RT-DETR](https://docs.ultralytics.com/models/rtdetr) which consume significantly more CUDA memory. ## Development and Deployment Experience @@ -118,17 +118,17 @@ YOLOX is a strong choice for: YOLOv8 is recommended for: -- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect/), [segmentation](https://docs.ultralytics.com/tasks/segment/), [classification](https://docs.ultralytics.com/tasks/classify/), and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within the Ultralytics ecosystem. +- **Versatile Multi-Task Deployment:** Projects requiring a proven model for [detection](https://docs.ultralytics.com/tasks/detect), [segmentation](https://docs.ultralytics.com/tasks/segment), [classification](https://docs.ultralytics.com/tasks/classify), and [pose estimation](https://docs.ultralytics.com/tasks/pose) within the Ultralytics ecosystem. - **Established Production Systems:** Existing production environments already built on the YOLOv8 architecture with stable, well-tested deployment pipelines. - **Broad Community and Ecosystem Support:** Applications benefiting from YOLOv8's extensive tutorials, third-party integrations, and active community resources. ### When to Choose Ultralytics (YOLO26) -For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26/) offers the best combination of performance and developer experience: +For most new projects, [Ultralytics YOLO26](https://docs.ultralytics.com/models/yolo26) offers the best combination of performance and developer experience: - **NMS-Free Edge Deployment:** Applications requiring consistent, low-latency inference without the complexity of Non-Maximum Suppression post-processing. - **CPU-Only Environments:** Devices without dedicated GPU acceleration, where YOLO26's up to 43% faster CPU inference provides a decisive advantage. -- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. +- **Small Object Detection:** Challenging scenarios like [aerial drone imagery](https://docs.ultralytics.com/datasets/detect/visdrone) or IoT sensor analysis where ProgLoss and STAL significantly boost accuracy on tiny objects. ## Looking Ahead: The YOLO26 Architecture @@ -136,12 +136,12 @@ While YOLOv8 provides exceptional balance and usability, the frontier of artific YOLO26 introduces an **end-to-end NMS-free design**, completely eliminating the heuristic non-maximum suppression post-processing step. This breakthrough ensures stable, deterministic latency across diverse deployment targets. Furthermore, by deliberately removing the Distribution Focal Loss (DFL) module, YOLO26 achieves up to **43% faster CPU inference**, making it the absolute best choice for embedded systems and mobile applications. -Training stability is also revolutionized in YOLO26 through the integration of the novel **MuSGD optimizer**—a hybrid of SGD and Muon that accelerates convergence. Coupled with the new **ProgLoss + STAL** loss functions, YOLO26 delivers notable improvements in small-object recognition, which is highly critical for drone mapping and [security alarm systems](https://docs.ultralytics.com/reference/solutions/security_alarm/). +Training stability is also revolutionized in YOLO26 through the integration of the novel **MuSGD optimizer**—a hybrid of SGD and Muon that accelerates convergence. Coupled with the new **ProgLoss + STAL** loss functions, YOLO26 delivers notable improvements in small-object recognition, which is highly critical for drone mapping and [security alarm systems](https://docs.ultralytics.com/reference/solutions/security_alarm). ## Conclusion and Recommendations When evaluating older frameworks against modern solutions, the trajectory is clear. While YOLOX was an instrumental stepping stone in the transition to anchor-free methodologies, its lack of an integrated, multi-task ecosystem limits its utility in fast-paced production environments. -For developers prioritizing a seamless experience, versatile task support, and strong community backing, [YOLOv8](https://docs.ultralytics.com/models/yolov8/) remains a highly robust choice. However, for those looking to maximize edge computing performance, eliminate NMS bottlenecks, and achieve the highest possible accuracy with the latest training innovations, **[YOLO26](https://docs.ultralytics.com/models/yolo26/)** is overwhelmingly the recommended model for any new computer vision project. +For developers prioritizing a seamless experience, versatile task support, and strong community backing, [YOLOv8](https://docs.ultralytics.com/models/yolov8) remains a highly robust choice. However, for those looking to maximize edge computing performance, eliminate NMS bottlenecks, and achieve the highest possible accuracy with the latest training innovations, **[YOLO26](https://docs.ultralytics.com/models/yolo26)** is overwhelmingly the recommended model for any new computer vision project. -If you are interested in exploring other models within the Ultralytics suite, you may also want to review the performance characteristics of [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or read up on the pioneering NMS-free concepts originally tested in [YOLOv10](https://docs.ultralytics.com/models/yolov10/). +If you are interested in exploring other models within the Ultralytics suite, you may also want to review the performance characteristics of [YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) or read up on the pioneering NMS-free concepts originally tested in [YOLOv10](https://docs.ultralytics.com/models/yolov10). diff --git a/docs/en/compare/yolox-vs-yolov9.md b/docs/en/compare/yolox-vs-yolov9.md index 4a6b8bfacd7..809e440ddbb 100644 --- a/docs/en/compare/yolox-vs-yolov9.md +++ b/docs/en/compare/yolox-vs-yolov9.md @@ -32,7 +32,7 @@ YOLOX introduced several key changes to the standard detection pipeline. It impl ### Strengths and Limitations -The primary strength of YOLOX lies in its simplified design. The anchor-free mechanism means developers spend less time running clustering algorithms to find optimal anchor sizes for their specific data. However, as an older architecture natively built without recent advancements in self-attention or gradient pathing, it struggles to match the parameter efficiency of newer networks. It also lacks native support for advanced tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment/) and [pose estimation](https://docs.ultralytics.com/tasks/pose/) within a unified API. +The primary strength of YOLOX lies in its simplified design. The anchor-free mechanism means developers spend less time running clustering algorithms to find optimal anchor sizes for their specific data. However, as an older architecture natively built without recent advancements in self-attention or gradient pathing, it struggles to match the parameter efficiency of newer networks. It also lacks native support for advanced tasks like [instance segmentation](https://docs.ultralytics.com/tasks/segment) and [pose estimation](https://docs.ultralytics.com/tasks/pose) within a unified API. [Learn more about YOLOX](https://github.com/Megvii-BaseDetection/YOLOX){ .md-button } @@ -45,7 +45,7 @@ Fast forward to 2024, YOLOv9 introduced a highly theoretical approach to solving - **Release Date:** February 21, 2024 - **Reference:** [Arxiv Paper](https://arxiv.org/abs/2402.13616) - **Source Code:** [YOLOv9 GitHub Repository](https://github.com/WongKinYiu/yolov9) -- **Documentation:** [Ultralytics YOLOv9 Docs](https://docs.ultralytics.com/models/yolov9/) +- **Documentation:** [Ultralytics YOLOv9 Docs](https://docs.ultralytics.com/models/yolov9) ### Architectural Innovations @@ -53,9 +53,9 @@ YOLOv9's defining feature is Programmable Gradient Information (PGI), which ensu ### Strengths and Limitations -YOLOv9 excels in pushing the theoretical limits of [model accuracy](https://docs.ultralytics.com/guides/yolo-performance-metrics/). It yields fantastic mAP scores on COCO, making it a favorite for researchers. However, despite its efficiency, YOLOv9 still relies on traditional Non-Maximum Suppression (NMS) for post-processing, which introduces latency spikes during inference. For engineers focused on deploying AI to [edge devices](https://docs.ultralytics.com/guides/model-deployment-options/), managing NMS logic adds unnecessary complexity to the deployment pipeline. +YOLOv9 excels in pushing the theoretical limits of [model accuracy](https://docs.ultralytics.com/guides/yolo-performance-metrics). It yields fantastic mAP scores on COCO, making it a favorite for researchers. However, despite its efficiency, YOLOv9 still relies on traditional Non-Maximum Suppression (NMS) for post-processing, which introduces latency spikes during inference. For engineers focused on deploying AI to [edge devices](https://docs.ultralytics.com/guides/model-deployment-options), managing NMS logic adds unnecessary complexity to the deployment pipeline. -[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9/){ .md-button } +[Learn more about YOLOv9](https://docs.ultralytics.com/models/yolov9){ .md-button } !!! tip "Post-Processing Bottlenecks" @@ -90,11 +90,11 @@ While evaluating historical models like YOLOX and YOLOv9 provides valuable conte YOLO26 completely solves the post-processing bottlenecks of its predecessors with a **native end-to-end NMS-free design**, ensuring simpler deployment across all hardware. Furthermore, by removing Distribution Focal Loss (DFL) and integrating the novel **MuSGD Optimizer**—a hybrid of Stochastic Gradient Descent and Muon—YOLO26 achieves unprecedented training stability. -For developers deploying to constrained environments like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi/), YOLO26 delivers up to **43% faster CPU inference**. It also introduces **ProgLoss + STAL** loss functions, resulting in dramatic improvements in small-object recognition, which is critical for [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone/) and drone analytics. +For developers deploying to constrained environments like the [Raspberry Pi](https://docs.ultralytics.com/guides/raspberry-pi), YOLO26 delivers up to **43% faster CPU inference**. It also introduces **ProgLoss + STAL** loss functions, resulting in dramatic improvements in small-object recognition, which is critical for [aerial imagery](https://docs.ultralytics.com/datasets/detect/visdrone) and drone analytics. ### Streamlined Development Ecosystem -Unlike standalone research repositories, the Ultralytics ecosystem provides an unparalleled developer experience. Utilizing the [Ultralytics Python API](https://docs.ultralytics.com/usage/python/), engineers can drastically reduce boilerplate code. Furthermore, memory requirements are kept highly optimized, meaning you can train robust models using less GPU VRAM compared to heavily attention-based architectures. +Unlike standalone research repositories, the Ultralytics ecosystem provides an unparalleled developer experience. Utilizing the [Ultralytics Python API](https://docs.ultralytics.com/usage/python), engineers can drastically reduce boilerplate code. Furthermore, memory requirements are kept highly optimized, meaning you can train robust models using less GPU VRAM compared to heavily attention-based architectures. ```python from ultralytics import YOLO @@ -109,7 +109,7 @@ results = model.train(data="coco8.yaml", epochs=100, imgsz=640) model.export(format="engine", half=True) # Exports to TensorRT ``` -Beyond detection, YOLO26 seamlessly supports a multitude of tasks within the exact same framework. Whether you need precise [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb/) for satellite imaging or fine-grained pixel masks for [medical imaging applications](https://docs.ultralytics.com/datasets/detect/brain-tumor/), the workflow remains identical. For teams invested in previous generation workflows, [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) is also available and fully supported. +Beyond detection, YOLO26 seamlessly supports a multitude of tasks within the exact same framework. Whether you need precise [Oriented Bounding Boxes (OBB)](https://docs.ultralytics.com/tasks/obb) for satellite imaging or fine-grained pixel masks for [medical imaging applications](https://docs.ultralytics.com/datasets/detect/brain-tumor), the workflow remains identical. For teams invested in previous generation workflows, [Ultralytics YOLO11](https://platform.ultralytics.com/ultralytics/yolo11) is also available and fully supported. ## Ideal Use Cases and Deployment Strategies @@ -117,7 +117,7 @@ Choosing the right architecture depends entirely on your target deployment envir ### Edge Computing and Robotics -For low-power devices, relying on models that require heavy post-processing can cripple performance. While YOLOX-Nano is incredibly small, its accuracy is often insufficient for safety-critical tasks. YOLO26 is the definitive choice here; its lack of DFL and NMS allows it to run smoothly on raw CPU threads, making it perfect for autonomous robotics or [smart parking management](https://docs.ultralytics.com/guides/parking-management/). +For low-power devices, relying on models that require heavy post-processing can cripple performance. While YOLOX-Nano is incredibly small, its accuracy is often insufficient for safety-critical tasks. YOLO26 is the definitive choice here; its lack of DFL and NMS allows it to run smoothly on raw CPU threads, making it perfect for autonomous robotics or [smart parking management](https://docs.ultralytics.com/guides/parking-management). ### Academic Benchmarking @@ -125,6 +125,6 @@ If the sole goal is analyzing gradient flow and studying deep network bottleneck ### Enterprise Video Analytics -For large-scale video processing tasks like [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system/) or traffic monitoring, speed and versatile export capabilities are paramount. The native export tools provided by the Ultralytics framework allow teams to compile YOLO26 directly to [TensorRT](https://docs.ultralytics.com/integrations/tensorrt/) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino/) in a single command, drastically reducing time-to-market. +For large-scale video processing tasks like [security alarm systems](https://docs.ultralytics.com/guides/security-alarm-system) or traffic monitoring, speed and versatile export capabilities are paramount. The native export tools provided by the Ultralytics framework allow teams to compile YOLO26 directly to [TensorRT](https://docs.ultralytics.com/integrations/tensorrt) or [OpenVINO](https://docs.ultralytics.com/integrations/openvino) in a single command, drastically reducing time-to-market. By leveraging the comprehensive features of the Ultralytics ecosystem, machine learning teams can bypass the complexities of raw research codebases and focus directly on building scalable, real-world AI applications.