SkaarFacee/xwod-dataset

XWOD v1.1 — Extreme Weather Object Detection Dataset

XWOD is an object detection dataset of 10,010 images collected from publicly available online videos, contributor dashcams, and public photographs under 7 extreme weather conditions, annotated with axis-aligned bounding boxes in YOLO format.

• Classes (YOLO indices 0–5): person, car, truck, motorcycle, bus, bike
• Weather conditions: dust, flooding, fog, heavy_rain, snow, tornado, wildfire
• Splits: official train / valid / test — use verbatim, do not re-split.
• Total annotations: 42930 in 10010 images.

Repository layout

images/train/    images/valid/    images/test/    (JPEG, 10,010)
labels/train/    labels/valid/    labels/test/    (YOLO .txt, 10,010)
data.yaml                         Ultralytics-style config (repo-root relative)
README.md, xwod_release_notes.md, stats/xwod_stats.json, eda/xwod_sample_grid.png

Splits

Split	Images	Annotations	Paths
train	5945	25494	images/train + labels/train
valid	1744	7629	images/valid + labels/valid
test	2321	9807	images/test + labels/test

Class distribution

Index	Class	Annotations	Share
0	person	10632	24.8%
1	car	23158	53.9%
2	truck	5224	12.2%
3	motorcycle	1744	4.1%
4	bus	633	1.5%
5	bike	1539	3.6%

Class imbalance (max/min): 36.6× globally (train 40.3x, valid 28.3x, test 71.3x).

Weather distribution

Weather	Images	Share
dust	560	5.6%
flooding	5151	51.5%
fog	306	3.1%
heavy_rain	665	6.6%
snow	1203	12.0%
tornado	1164	11.6%
wildfire	961	9.6%

Every split covers all 7 conditions; weather is encoded in the filename prefix (<weather>_<split>_<seq>).

Example

License, provenance, and redistribution

Copyright remains with the respective copyright owners. XWOD is released exclusively for non-commercial academic research. Redistribution or commercial use of the complete dataset is not permitted without authorization.

XWOD is released exclusively for non-commercial academic research, scientific benchmarking, and educational use. Please cite the XWOD paper when using the dataset. Redistribution or commercial use of the complete dataset is not permitted without authorization; this private repository serves the group only.

For copyright, privacy, or takedown requests (name, contact email, image ID, description, supporting evidence): t111c71102@ntut.org.tw.

Protocol

Training/evaluation protocol and conversion recipes: see docs/xwod-protocol.md in the source workspace (or the stats/ JSON for full statistics).

Citation

@misc{xwod2026,
  title={XWOD: An Extreme Weather Object Detection Dataset},
  author={XWOD contributors},
  year={2026},
  note={v1.1, non-commercial academic use}
}