UAVDT: UAV Detection and Tracking Benchmark (No Data Mirror -- Adapter Only)

This card describes a DetectionBench dataset adapter for UAVDT. It does NOT host or redistribute the dataset itself -- UAVDT is distributed "for research purpose only" with no redistribution grant. See Getting the Data for the official download links.

Disclaimer

DetectionBench is not an official release of UAVDT and does not host any UAVDT images, annotations, or derived files anywhere -- not on Hugging Face, not in this repository.

UAVDT was created by Dawei Du, Yuankai Qi, Hongyang Yu, Yifan Yang, Kaiwen Duan, Guorong Li, Weigang Zhang, Qingming Huang, and Qi Tian, who retain all rights. Its official page states the dataset "is for research purpose only" -- a stated restriction, not merely an unclear one, so DetectionBench does not attempt to reinterpret it as permitting redistribution.

What this repository provides instead:

1. A DetectionBench dataset adapter that converts an official UAVDT download into DetectionBench's canonical training layout, once you have obtained the data yourself.
2. A helper command, detectionbench-download-dataset --dataset uavdt, that prints the official download links (and can fetch the Google Drive ones automatically via gdown) -- see Getting the Data.
3. This banner, generated locally from a UAVDT copy already converted through the adapter, purely to illustrate the dataset's domain and annotation style.

---

Dataset Description

UAVDT is an aerial vehicle-detection and -tracking benchmark: roughly 80,000 frames drawn from ~10 hours of raw UAV video, annotated for car / truck / bus detection with additional per-object attributes (occlusion level, out-of-view fraction) and per-sequence conditions (weather, altitude, camera viewpoint). The official split is 30 training sequences (24,143 frames) and 70 test sequences (53,676 frames), with no official validation split.

---

Getting the Data

UAVDT is not mirrored here. Get it directly from the authors:

pip install "detectionbench[download]"   # installs gdown, for the Google Drive options
detectionbench-download-dataset --dataset uavdt

This prints every official download option (and attempts the primary Google Drive link automatically). The two you need for detection:

Resource	Link
UAVDT-Benchmark-M (frames + annotations)	https://drive.google.com/file/d/1m8KA6oPIRK_Iwt9TYFquC87vBc_8wRVc/view
DET/MOT toolkit (optional, official eval scripts)	https://drive.google.com/open?id=19498uJd7T9w4quwnQEy62nibt3uyT9pq

(UAVDT-Benchmark-S is single-object-tracking data, not needed for detection.)

Once downloaded, convert it into DetectionBench's canonical layout:

detectionbench-prepare-coco --dataset uavdt --raw-dir <downloaded_dir> --output-dir <coco_out>
detectionbench-convert-coco-to-yolo --input-dir <coco_out> --output-dir <yolo_out>

The adapter keeps UAVDT's official test split and carves a sequence-aware validation set out of train (6 of 30 sequences held out, seeded -- verified disjoint from the remaining 24), so no video frame leaks across the train/val boundary.

Classes (3)

car, truck, bus -- the canonical UAVDT-DET taxonomy. (The Dataset Ninja Supervisely re-export this adapter also accepts additionally lists a vehicle class that is unused in practice.)

---

Dataset Sources

Original Paper

The Unmanned Aerial Vehicle Benchmark: Object Detection and Tracking

Dawei Du, Yuankai Qi, Hongyang Yu, Yifan Yang, Kaiwen Duan, Guorong Li, Weigang Zhang, Qingming Huang, Qi Tian

European Conference on Computer Vision (ECCV), 2018. arXiv:1804.00518

Official Resources

• Homepage: https://sites.google.com/view/grli-uavdt

---

License

"For research purpose only" -- stated directly on the official homepage. This is a real, explicit restriction, not merely an undocumented one, so:

• No Hugging Face mirror of the data is provided or planned.
• The DetectionBench adapter is provided for local, research use against a copy you download yourself and agree to those terms for.
• If you need broader rights (redistribution, commercial use), contact the original authors.

---

Citation

If you use this dataset, please cite:

@InProceedings{du2018unmanned,
  title={The Unmanned Aerial Vehicle Benchmark: Object Detection and Tracking},
  author={Du, Dawei and Qi, Yuankai and Yu, Hongyang and Yang, Yifan and Duan, Kaiwen and Li, Guorong and Zhang, Weigang and Huang, Qingming and Tian, Qi},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2018}
}

---

Acknowledgements

We sincerely thank Dawei Du, Yuankai Qi, Hongyang Yu, and their co-authors for creating and publicly releasing this valuable aerial vehicle-detection benchmark.