FOTBCD-CC

FOTBCD-CC adds natural-language change descriptions to the before-and-after aerial image pairs in FOTBCD. It contains 321,200 English captions for 223,675 temporal pairs covering 28 departments across mainland France.

This repository contains the caption annotations only. The corresponding images remain in FOTBCD and are linked through a pinned dataset revision, shard path, and WebDataset sample key. The imagery is not duplicated here.

Dataset summary

Split	Temporal pairs	Captions	Durable change	No durable change
Train	221,712	318,162	132,720	88,992
Validation	984	1,532	704	280
Test	979	1,506	692	287
Total	223,675	321,200	134,116	89,559

Every pair has between one and five captions. Captions average 26.13 words.

Annotation format

Each row represents one before-and-after pair:

Field	Description
sample_id	Globally unique split and sample identifier
split	train, val, or test
image_id	Image-pair ID from FOTBCD
sample_key	Key used inside the FOTBCD WebDataset shard
image_shard	Path of the corresponding TAR shard in FOTBCD
before_member / after_member	Image member names inside that shard
department	French department code
before_year / after_year	Acquisition years
epsg	Spatial reference identifier; all samples use EPSG:2154
crop_stratum	Sampling stratum used during caption generation
patch_category	Patch category from FOTBCD
building_annotation_count	Number of FOTBCD building instances in the pair
durable_change	Whether the captioner identified a lasting physical change
captions	One to five generated descriptions
caption_count	Number of descriptions in captions

The published Parquet files intentionally exclude inference-only logs such as latency, token usage, retries, duplicated raw responses, and empty error fields.

Durable change

A durable change is a lasting physical alteration visible between the two dates, such as construction, demolition, a new paved road, an expanded building, or cleared woodland converted to another use.

Differences caused only by shadows, parked vehicles, seasonal vegetation, cultivation cycles, or surface appearance are treated as no durable change. No-change rows still contain a scene description.

Load the captions

from datasets import load_dataset

dataset = load_dataset("abdevd/FOTBCD-CC")
row = dataset["train"][0]

print(row["durable_change"])
print(row["captions"])
print(row["image_shard"], row["sample_key"])

Join captions with the FOTBCD images

The Parquet rows point to the exact WebDataset shards and members in the pinned FOTBCD revision:

from datasets import load_dataset
import webdataset as wds

BASE_REVISION = "92b677c881f0d973d76460204431c711ccbac25b"
BASE_URL = (
    "https://huggingface.co/datasets/abdevd/FOTBCD/resolve/"
    f"{BASE_REVISION}/"
)

captions = load_dataset("abdevd/FOTBCD-CC", split="train")
caption_by_key = {row["sample_key"]: row for row in captions}
shards = sorted({BASE_URL + row["image_shard"] for row in captions})

images = wds.WebDataset(shards, shardshuffle=True).decode("pil")

for sample in images:
    annotation = caption_by_key[sample["__key__"]]
    before_image = sample["before.png"]
    after_image = sample["after.png"]
    pair_captions = annotation["captions"]
    # Train or evaluate your model here.

Install the required libraries with:

pip install datasets webdataset pillow

Generation process

The full annotations were generated by a Qwen3-VL-8B model adapted from Qwen3.8-Max teacher outputs.

Qwen3.8-Max first annotated a selected teacher set of 19,997 temporal pairs, producing 29,447 captions. Those responses and durable-change decisions were then used for supervised LoRA response distillation into Qwen3-VL-8B. During fine-tuning, the vision tower and visual aligner were frozen while the language model was adapted. The resulting model was served through vLLM to annotate all 223,675 FOTBCD pairs.

The student examines the complete before-and-after crops, predicts whether a durable change occurred, and produces one or more descriptions of the visible events or stable scene. Global annotation provenance, including the prompt hash, is recorded in provenance.json rather than repeated in every Parquet row.

The source JSONL contained one successful, valid response for every FOTBCD pair. Before export, all 223,675 split/image identifiers and image paths were checked against the FOTBCD manifests. No missing pairs, duplicate keys, path mismatches, invalid responses, empty caption lists, or blank captions were found.

Licensing

The FOTBCD-CC caption annotations are licensed under CC BY 4.0.

The corresponding imagery and building data are provided by the Institut national de l’information géographique et forestière through BD ORTHO and BD TOPO, under the Etalab Open Licence 2.0. See the FOTBCD dataset card for the base dataset's documentation and attribution.

Citation

When using the paired FOTBCD imagery, please cite:

@article{moubane2026fotbcd,
  title   = {FOTBCD: A Large-Scale Building Change Detection Benchmark from French Orthophotos and Topographic Data},
  author  = {Moubane, Abdelrrahman},
  journal = {arXiv preprint arXiv:2601.22596},
  year    = {2026},
  url     = {https://arxiv.org/abs/2601.22596}
}