Testing change detection beyond one city
Building-change models are often trained and tested on imagery from a narrow geographic area. LEVIR-CD+ covers regions in Texas, while WHU-CD comes from one image pair over Christchurch. A model can score well in-domain while learning building styles, urban layouts, vegetation, and imaging conditions that do not transfer elsewhere.
I built FOTBCD to make geographic domain shift part of the benchmark rather than an afterthought. The full dataset is now openly available on Hugging Face: 223,675 georeferenced before-and-after pairs, or 447,350 images, with 960,880 COCO-format building instances across 28 departments in mainland France. Twenty-five departments supply training data; three geographically separate departments are reserved for validation and testing, with no overlap between their locations and the training geography.
The coverage includes dense cities, suburbs, agricultural regions, coasts, industrial areas, and mountainous terrain. All imagery is supplied at 0.2 metres per pixel, so the benchmark varies geography without reducing buildings to coarse satellite-scale shapes.
Building the pairs from authoritative French data
FOTBCD combines two datasets maintained by IGN, France’s national mapping agency:
- BD ORTHO provides orthorectified RGB aerial imagery from repeated national surveys, generally separated by three to four years.
- BD TOPO provides authoritative building footprints and attributes with metric precision.
I aligned successive BD TOPO snapshots with the corresponding BD ORTHO acquisitions, inferred building-level transitions, and converted them into image pairs and masks. Each pair is a 512 × 512 crop with its spatial metadata retained in Lambert-93 (EPSG:2154).
The full release makes the complete instance-annotated collection available alongside the two benchmark subsets introduced with the paper:
| Dataset | Coverage | Pairs | Annotation |
|---|---|---|---|
| FOTBCD | 28 departments | 223,675 | 960,880 COCO polygons labelled UNCHANGED, DEMOLISHED, or NEW |
| FOTBCD-Binary | 28 departments | 27,871 | Pixel-wise building-change mask |
| FOTBCD-Instances | 6 departments | 4,000 | Instance-annotated subset released with the paper |
The complete FOTBCD release and FOTBCD-Instances preserve the direction of each transition and the identity of individual buildings, supporting instance-aware experiments and boundary-level analysis. FOTBCD-Binary merges new and demolished buildings into one change mask for dense segmentation and retains the compact benchmark used for the paper’s cross-dataset experiments.
Splitting by geography, not random patches
A random patch split would leak local appearance into every partition. Instead, the binary benchmark holds out three complete departments. Training labels are generated at scale from temporal topographic differences using rule-based filtering, domain heuristics, learned checks, and the aligned aerial imagery. The held-out validation and test samples are then manually selected and verified.
Quality control covers acquisition-date consistency, building geometry and metadata, implausible transitions, residual edge cases, and final human review. This design accepts that a large automatically generated training set can contain some noise while keeping the evaluation sets reliable and geographically independent.
A fixed baseline for cross-dataset evaluation
I used the same HybridSiam-CD configuration for FOTBCD-Binary, LEVIR-CD+, and WHU-CD so the comparison measured the training data rather than changes in model capacity.
HybridSiam-CD combines a frozen DINOv3 SAT-493M vision transformer, which extracts semantic features from both dates, with a siamese ImageNet-pretrained ResNet34 for multi-scale spatial and boundary detail. A lightweight decoder fuses their differences and skip features into a dense change mask. Each model was trained for 50,000 steps under the same optimizer, schedule, augmentation, and loss configuration.
The cross-dataset IoU matrix was:
| Train dataset ↓ / Test dataset → | FOTBCD-Binary | LEVIR-CD+ | WHU-CD |
|---|---|---|---|
| FOTBCD-Binary | 0.8180 | 0.2984 | 0.6974 |
| LEVIR-CD+ | 0.3003 | 0.7365 | 0.5432 |
| WHU-CD | 0.3417 | 0.2127 | 0.8939 |
The diagonal scores show that every dataset is easiest in-domain. The transfer results expose the more important pattern: models trained on LEVIR-CD+ or WHU-CD reached only 0.3003 and 0.3417 IoU on FOTBCD-Binary, while the FOTBCD-trained model reached 0.6974 on WHU-CD. Transfer to LEVIR-CD+ remained difficult in both directions. The result is not that geographic diversity removes domain shift, but that it produces more balanced transfer than training on one constrained region.
What I released
The open Hugging Face release contains 171 WebDataset TAR shards totalling about 170 GiB. Each sample groups the earlier image, later image, and a JSON record containing its spatial metadata and COCO polygon annotations. The training split contains 221,712 pairs; the geographically held-out validation and test splits contain 984 and 979 pairs.
The full dataset is licensed under CC BY 4.0. Its source imagery and building data come from IGN’s BD ORTHO and BD TOPO products under the Etalab Open Licence 2.0. The project repository continues to provide the paper’s binary and instance subsets, training and evaluation code, configuration files, and pretrained weights.
The release is designed for streaming, so experiments can read the shards directly from Hugging Face without first downloading the complete archive:
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")
)The same broader image collection later became the foundation for FOTBCD-CC, where I moved beyond building masks to open-ended descriptions and retrieval across whole scenes.
Limitations
FOTBCD covers varied environments, but they are all in mainland France; international transfer still requires separate evaluation. Its annotations concern buildings, not roads, vegetation, or land-use change. The binary benchmark merges construction and demolition into one class. The full release preserves those classes separately, but its training annotations were generated at scale and can contain label noise, imperfect geometry, or missed transitions. The class distribution is also imbalanced, especially for demolished buildings.