[![Hugging Face](https://img.shields.io/badge/🤗%20Hugging%20Face-Full%20FOTBCD-yellow)](https://huggingface.co/datasets/abdevd/FOTBCD) [![FOTBCD-Binary](https://img.shields.io/badge/🤗%20Hugging%20Face-FOTBCD--Binary-yellow)](https://huggingface.co/datasets/abdevd/FOTBCD-Binary) [![arXiv](https://img.shields.io/badge/arXiv-2601.22596-b31b1b.svg)](https://arxiv.org/abs/2601.22596) [![Code License: MIT](https://img.shields.io/badge/Code-MIT-green.svg)](LICENSE) [![Dataset License: CC BY 4.0](https://img.shields.io/badge/Datasets-CC%20BY%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by/4.0/) # FOTBCD: A Geographically Diverse Dataset for Building Change Detection from High-Resolution Aerial Imagery A large-scale building change detection benchmark from French orthophotos and topographic data. The full dataset is now openly available on [Hugging Face](https://huggingface.co/datasets/abdevd/FOTBCD).

## Datasets We openly release the complete FOTBCD collection alongside a compact binary benchmark: | Dataset | Departments | Pairs | Patch Size | Resolution | Annotation | License | |---------|-------------|-------|------------|------------|------------|---------| | **FOTBCD** | 28 (25 train / 3 eval) | 223,675 | 512x512 | 0.2m | 960,880 COCO polygons (`UNCHANGED`, `DEMOLISHED`, `NEW`) | CC BY 4.0 | | **FOTBCD-Binary** | 28 (25 train / 3 eval) | 27,871 | 512x512 | 0.2m | Binary mask | CC BY 4.0 | ---


Geographic coverage: 25 training departments (green), 3 held-out evaluation departments (blue)

### Download and streaming - [Full FOTBCD on Hugging Face](https://huggingface.co/datasets/abdevd/FOTBCD) - 223,675 pairs in 171 WebDataset TAR shards (about 170 GiB). - [FOTBCD-Binary on Hugging Face](https://huggingface.co/datasets/abdevd/FOTBCD-Binary) - 27,871 pairs with pixel-wise change masks. - [Google Drive](https://drive.google.com/drive/folders/1XpV4ouhvVg0M28S7-u7PbDhYA56pQHRm?usp=sharing) - FOTBCD-Binary and pretrained weights. The full release can be streamed directly without downloading the complete archive: ```bash pip install webdataset pillow ``` ```python import webdataset as wds url = ( "pipe:curl -s -L " "https://huggingface.co/datasets/abdevd/FOTBCD/resolve/main/" "data/train/train-{000000..000168}.tar" ) dataset = ( wds.WebDataset(url, shardshuffle=True) .shuffle(5_000) .decode("pil") .to_tuple("before.png", "after.png", "json") ) before_image, after_image, annotation = next(iter(dataset)) ``` ### Structure ``` FOTBCD/ data/ train/ train-000000.tar ... train-000168.tar val/ val-000000.tar test/ test-000000.tar # Every WebDataset sample contains: {id}.before.png {id}.after.png {id}.json # spatial metadata and COCO polygon annotations FOTBCD-Binary/ images/ train/ before/ # {id}.png after/ # {id}.png label/ # binary mask (0=no change, 255=change) val/ test/ ``` ## Installation ```bash conda env create -f environment.yml conda activate fotbcd ``` Install PyTorch separately based on your system: https://pytorch.org/get-started/locally/ Set the dataset paths in `config.py` before training. ## Training ```bash python train.py ``` ### Cross-Dataset Evaluation ```bash python evaluate.py --checkpoints_dir ./checkpoints --batch_size 32 ``` ## Results Cross-domain generalization (IoU): | ↓ Train / Test → | FOTBCD-Binary | LEVIR-CD+ | WHU-CD | |--------------|--------|-----------|--------| | FOTBCD-Binary| 0.818 | 0.299 | 0.697 | | LEVIR-CD+ | 0.300 | 0.737 | 0.544 | | WHU-CD | 0.342 | 0.213 | 0.894 | Cross-domain generalization (F1): | ↓ Train / Test → | FOTBCD-Binary | LEVIR-CD+ | WHU-CD | |--------------|--------|-----------|--------| | FOTBCD-Binary| 0.900 | 0.460 | 0.822 | | LEVIR-CD+ | 0.462 | 0.848 | 0.704 | | WHU-CD | 0.509 | 0.351 | 0.944 | Cross-domain generalization (Precision): | ↓ Train / Test → | FOTBCD-Binary | LEVIR-CD+ | WHU-CD | |--------------|--------|-----------|--------| | FOTBCD-Binary| 0.915 | 0.819 | 0.803 | | LEVIR-CD+ | 0.802 | 0.880 | 0.829 | | WHU-CD | 0.736 | 0.821 | 0.956 | Cross-domain generalization (Recall): | ↓ Train / Test → | FOTBCD-Binary | LEVIR-CD+ | WHU-CD | |--------------|--------|-----------|--------| | FOTBCD-Binary| 0.886 | 0.320 | 0.841 | | LEVIR-CD+ | 0.324 | 0.819 | 0.612 | | WHU-CD | 0.390 | 0.223 | 0.933 | ## License - **Code**: MIT - **Full FOTBCD dataset**: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) - **FOTBCD-Binary**: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) - **Source Data**: BD ORTHO / BD TOPO by IGN under [Licence Ouverte 2.0](https://alliance.numerique.gouv.fr/licence-ouverte-open-licence/) ## Citation If you use FOTBCD in your research, please cite our paper: ```bibtex @misc{moubane2026fotbcd, title={FOTBCD: A Large-Scale Building Change Detection Benchmark from French Orthophotos and Topographic Data}, author={Abdelrrahman Moubane}, year={2026}, eprint={2601.22596}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2601.22596}, } ``` ## Acknowledgments FOTBCD is derived from BD ORTHO and BD TOPO databases by [Institut national de l'information géographique et forestière (IGN)](https://www.ign.fr/).