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.

Datasets

We release two community versions derived from FOTBCD, both freely available under CC BY-NC-SA 4.0:

Dataset	Departments	Pairs	Patch Size	Resolution	Annotation	License
FOTBCD-Binary	28 (25 train / 3 eval)	~28k	512x512	0.2m	Binary mask	CC BY-NC-SA 4.0
FOTBCD-Instances	6 (3 train / 3 eval)	4k	512x512	0.2m	COCO polygons	CC BY-NC-SA 4.0

---

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

Download

Download (Google Drive) - contains FOTBCD-Binary, FOTBCD-Instances, and pretrained weights.

Structure

FOTBCD-Binary/
    images/
        train/
            before/     # {id}.png
            after/      # {id}.png
            label/      # binary mask (0=no change, 255=change)
        val/
        test/

FOTBCD-Instances/
    images/
        train/
            before/
            after/
        val/
        test/
    annotations/
        train.json      # COCO format (categories: UNCHANGED=1, DEMOLISHED=2, NEW=3)
        val.json
        test.json

Installation

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

python train.py

Cross-Dataset Evaluation

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
• FOTBCD-Binary / FOTBCD-Instances: CC BY-NC-SA 4.0
• Source Data: BD ORTHO / BD TOPO by IGN under Licence Ouverte 2.0

FOTBCD (Commercial License)

For industrial applications requiring larger scale and full instance-level annotations:

220,000+ image pairs	Multi-region coverage across France
950k+ building polygons	NEW / DEMOLISHED / UNCHANGED per instance
Commercial license	For production deployment and proprietary applications

Contact info@retgen.ai for licensing inquiries. Academic collaborations are welcome.

Citation

If you use FOTBCD in your research, please cite our paper:

@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).