Carousel is a dataset of 277 images containing multiple subjects. We present it for the pursuit of performing multi-target automatic image cropping to derive multiple crops from a single photograph, with minimal user input.
Each image contains two or more salient regions and has an average resolution of 10.58 MP. Carousel is organized into two directories with the following structures and file formats:
This directory contains the images in their original form. Each photograph has two sidecar files with the naming structure shown below:
└─ dataset
└─ image.jpeg
└─ image.json
└─ image_metadata.json
image.json would contain the label coordinates, along with the image's height and width and other information.
image_metadata.json would contain the following information:
website_url- A URL directly to the image.subcrops- The number of crops that should be derived from the image, which also indicates the number of salient regions.title- The title of the image as given by where it was sourced from.creator- Creator of the image.copyright_license_type- The type of copyright the image falls under.
This directory contains the images divided into distinct subregions.
└─ partitions
└─ image
└─ image_{aspect_ratio}_partition_1.jpg
└─ ...
└─ image_{aspect_ratio}_partition_{k}.jpg
└─ partition_info.json
Each image will have partitions according to the value defined in the subcrops field in the _metadata sidecar file. They will also have a partition_info.json file, which records:
original_filename: File name for source image.aspect_ratio_chosen: Aspect ratio used in the saliency partitioning algorithm, either3:2or2:3.num_partitions: A reflection of thesubcropsfield.orientation: The orientation of the partitions (i.e.,verticalwould be partitioned along the x axis, v.v.).partition_files: the names of each partition generated for the source image.
This file lists images that were not considered successful during the partitioning process. These images usually feature partitions with extreme aspect ratios or that bisect salient regions.
Images were sourced from Wikimedia Commons and image aggregators such as Flickr and Rawpixel. All images were distributed with open, non-commercial licenses.