AyaHassan/fair-bridges-chips

fAIr Bridges training chips (phase 1)

256 × 256 px aerial image chips (Web Mercator zoom 18, about 0.6 m per pixel, the way fAIr cuts chips) with rotated bridge boxes for the fAIr Bridges base model (YOLO26s-OBB, one class: bridge).

What is inside

• data/shard-*.parquet: one row per chip. Columns include image (PNG), obb_yolo (YOLO-OBB lines 0 x1 y1 x2 y2 x3 y3 x4 y4, normalised 0–1), bridge_ids, the tile x, y, z, bounds_3857, the source image (image_url, licence_img, provider), and kind (positive, hard_negative, random_negative).
• index/chips-*.parquet: the same rows without images (fast analysis).
• index/bridges-*.parquet: one row per bridge, with its geometry as delivered (source_wkt: OSM way or outline, BD TOPO feature), the deck polygon used for the boxes (deck_wkt), size, direction and tags.
• splits/split.parquet: chip_id → train / val / test / buffer / drop, with the reason.
• MANIFEST.csv (sha256 of every file) and dataset_record.md (sources, licences, counts, tests).

Chips per split

source	test	train	val
france	0	1503	345
oam_osm	110	662	88

split	hard_negative	positive	random_negative
test	13	80	17
train	554	1090	521
val	84	242	107

How the split avoids leakage

Test = whole OAM regions (oam_osm/dar_es_salaam). Train / validation = ~2.4 km blocks (zoom-14 tiles) chosen by a stable hash; every chip of a bridge stays in one split; chips within 1 chip of another split are dropped; near-duplicate chips across splits are dropped. Leakage tests: all passed.

Licences and attribution

• Labels: © OpenStreetMap contributors, ODbL 1.0 (OAM + OSM chips); IGN BD TOPO, Etalab 2.0 (France chips). Because the database includes OSM-derived data, the dataset as a whole is released under the ODbL.
• Imagery: OpenAerialMap images keep their own open licence (listed per chip in licence_img, CC BY 4.0 family; non-commercial images are excluded) and must be credited to their providers (provider, image_url); IGN BD ORTHO, Etalab 2.0.

Load

from datasets import load_dataset
ds = load_dataset("AyaHassan/fair-bridges-chips", split="train", streaming=True)   # all shards; use splits/split.parquet to filter

To export a YOLO-OBB folder for training: python -m bridgebuild to-yolo --config <config> (see the repo README).