Constructing Urban Building Exteriors Databases with Street View Imagery and Multimodal Large Language Model
Official implementation for constructing urban building exterior databases from OpenStreetMap, Google Street View imagery, and GPT-4o-based multimodal annotation.
Paper | Installation | Workflow | Usage | Citation
- 2025-11-12: BuildingView first appeared online as a SpatialDI 2025 conference paper in Lecture Notes in Computer Science, volume 15838.
- 2025: The paper was published in Spatial Data and Intelligence by Springer, Singapore.
Urban building exteriors are important for urban analytics, energy efficiency, environmental sustainability, architectural design, and human-centered planning. However, building exterior information is difficult to collect at scale and is often missing from conventional geospatial datasets.
BuildingView builds a reproducible workflow for constructing building exterior databases by integrating:
- OpenStreetMap building footprints and spatial metadata through the Overpass API
- Google Street View imagery for building-level exterior observation
- GPT-4o multimodal annotation guided by user-defined building exterior prompts
- GIS export and mapping for downstream urban analysis
The repository is aligned with the Springer chapter:
BuildingView: Constructing Urban Building Exteriors Databases with Street View Imagery and Multimodal Large Language Model
Zongrong Li, Yunlei Su, Hongrong Wang, Wufan Zhao
In: Spatial Data and Intelligence, SpatialDI 2025, Lecture Notes in Computer Science, vol. 15838, Springer, Singapore
DOI: 10.1007/978-981-95-3102-8_1
- Street-view-based exterior database construction: Samples building locations and retrieves corresponding Google Street View imagery.
- OSM-integrated spatial workflow: Uses Nominatim and Overpass API to query building footprints, coordinates, addresses, and building attributes.
- Prompt-driven GPT-4o annotation: Supports customized prompts for extracting building exterior indicators from street-view images.
- Validated urban case studies: The paper reports validation using data from New York City, Amsterdam, and Singapore.
- GIS-ready outputs: Exports annotated building records to CSV, GeoJSON, and Shapefile.
City / country query or bounding box
|
|-- Retrieve OSM building data Overpass.py / Overpass_bounding_box.py
|
|-- Optional sampled-location map map.py
|
|-- Download Google Street View images StreetView_donloader.py
|
|-- Annotate images with GPT-4o image_processing_pipeline.py / openai.py
|
|-- Merge annotations with source records image_processing_pipeline.py
|
`-- Export GIS files export_results.py
BuildingView/
|-- Overpass.py # Retrieve OSM building data by city and country
|-- Overpass_bounding_box.py # Retrieve OSM building data by bounding box
|-- city_country_matcher.py # Match city queries to city/country names
|-- StreetView_donloader.py # Download Google Street View images
|-- image_processing_pipeline.py # Run annotation, retry failures, and merge JSONL files
|-- openai.py # GPT-4o image annotation script
|-- map.py # Generate sampled-location maps
|-- export_results.py # Export CSV, GeoJSON, and Shapefile outputs
|-- Buildingview.png # Workflow figure
|-- prompt.txt # User-editable annotation prompt
|-- openai_api_keys.txt # OpenAI API keys, one per line
`-- requirements.txt
git clone https://github.com/Jasper0122/BuildingView
cd BuildingView
conda create -n buildingview python=3.9 -y
conda activate buildingview
pip install -r requirements.txtBuildingView requires:
- Google Street View Static API
- OpenAI API
- Internet access to Nominatim and Overpass API
Place OpenAI API keys in openai_api_keys.txt, one key per line:
sk-...
sk-...
Edit prompt.txt to define the building exterior indicators and output schema you want GPT-4o to extract.
Search candidate city/country names:
python city_country_matcher.py "NY"Retrieve buildings by city and country:
python Overpass.py "New York" "United States" 1000Retrieve buildings by bounding box:
python Overpass_bounding_box.py "New York" 1000 40.477399 -74.259090 40.917577 -73.700272python map.py "Data/New_York_United_States_1000.jsonl"python StreetView_donloader.py "Data/New_York_United_States_1000.jsonl" "YOUR_GOOGLE_API_KEY"Images are saved in a directory named after the input JSONL file.
python image_processing_pipeline.py \
"GoogleStreetViewImages/New_York_United_States_1000" \
"prompt.txt" \
"openai_api_keys.txt"The pipeline:
- Runs
openai.pyon street-view images - Logs failed images
- Retries failed requests
- Merges labels back into the source JSONL records
- Saves merged results under
result/
python export_results.py "result/New_York_United_States_1000.jsonl"The script exports:
- CSV
- GeoJSON
- Shapefile
BuildingView produces:
- Building-level JSONL records from OSM and Overpass
- Google Street View image collections
- GPT-4o building exterior annotations
- Merged building exterior databases
- GIS-ready exports for mapping and analysis
BuildingView focuses on street-view-based building exterior annotation. The follow-up project BuildingMultiView extends this idea to multi-perspective imagery by combining satellite house, satellite neighborhood, and street-view branches for multi-scale building characterization.
If you find this repository useful, please cite:
@inproceedings{li2026buildingview,
title = {BuildingView: Constructing Urban Building Exteriors Databases with Street View Imagery and Multimodal Large Language Model},
author = {Li, Zongrong and Su, Yunlei and Wang, Hongrong and Zhao, Wufan},
booktitle = {Spatial Data and Intelligence},
series = {Lecture Notes in Computer Science},
volume = {15838},
pages = {1--19},
publisher = {Springer},
address = {Singapore},
year = {2026},
doi = {10.1007/978-981-95-3102-8_1}
}For questions about the code or paper, please contact Wufan Zhao at wufanzhao@hkust-gz.edu.cn.
This project is released under the MIT License.