VisDrone2019-DET

VisDrone2019-DET, the drone-view object detection benchmark, packed once with the official annotations kept whole, so it loads in one line and no data path has to be configured:

from datasets import load_dataset

ds = load_dataset("shijli/visdrone2019-det")   # 6471 train / 548 validation / 1610 test

Splits

split	images	boxes	official name
train	6471	353550	VisDrone2019-DET-train
validation	548	40169	VisDrone2019-DET-val
test	1610	77547	VisDrone2019-DET-test-dev

The test split is the official test-dev set, whose labels VisDrone released. The test-challenge set has no public labels and is not included.

id is the original VisDrone file stem (e.g. 0000002_00005_d_0000014), so rows join trivially to a local VisDrone installation.

Columns

column	type	meaning
id	string	the VisDrone file stem
image	image	the JPEG, bytes unchanged
width, height	int32	image size in pixels
objects.category	ClassLabel(12)	the official category id, see below
objects.bbox	4 x int32	the official [x, y, w, h], absolute pixels, top-left origin
objects.score	int8	the official flag: 1 a real object, 0 an ignored box
objects.truncation	int8	0 none, 1 partly outside the frame
objects.occlusion	int8	0 none, 1 partial (1-50%), 2 heavy (50-100%)

objects is a dict of parallel lists, the way datasets stores a sequence of structs:

r = ds["train"][0]
for category, box in zip(r["objects"]["category"], r["objects"]["bbox"]):
    ...

Nothing is filtered: the annotations are the official eight columns, verbatim

Every VisDrone-DET mirror on the Hub has thrown something away. The YOLO conversions drop the ignored regions and the score / truncation / occlusion columns and renumber the classes; the COCO conversions drop the ignored regions and the others class and densify the ids. Which boxes to train on and which to ignore at evaluation time is a decision every codebase makes differently, so this packing makes none of it. Each split holds exactly as many boxes as the official annotation files, including the three zero-area boxes in the train set, and every box carries all eight official fields.

The category ClassLabel is indexed by the official VisDrone id, so 0 is still ignored regions and 11 is still others:

id	name	id	name
0	ignored regions	6	truck
1	pedestrian	7	tricycle
2	people	8	awning-tricycle
3	bicycle	9	bus
4	car	10	motor
5	van	11	others

Two of these are pseudo-classes, not objects. Ignored regions (0) are areas the annotators declared unscorable; the official VisDrone toolkit removes detections that fall inside them before scoring, and they should not be trained on as a class. Others (11) is the catch-all that the standard 10-class protocol leaves out. In every split, score == 0 holds for exactly the boxes of category 0 or 11, and for no others.

The usual 10-class protocol is therefore one filter away:

def keep_objects(objects):
    keep = [c not in (0, 11) for c in objects["category"]]
    return {k: [v for v, ok in zip(vals, keep) if ok] for k, vals in objects.items()}

and the ignored regions are still there, as category 0, for an evaluator that wants to mask them out.

How this was packed

The repository includes create_dataset.py, which produced these parquet files from the three official VisDrone2019-DET-* directories. Images are copied through byte for byte. Before pushing, the build was checked against the source: every split holds exactly as many boxes as its annotation files, and sampled records decode back to the same columns as the file they came from.

License and attribution

VisDrone is collected and annotated by the AISKYEYE team at the Lab of Machine Learning and Data Mining, Tianjin University, and distributed under CC BY-NC-SA 3.0 for non-commercial research use. This packaging adds no rights: use it under those same terms and cite the original work.

@article{zhu2021detection,
    title={Detection and Tracking Meet Drones Challenge},
    author={Zhu, Pengfei and Wen, Longyin and Du, Dawei and Bian, Xiao and Fan, Heng and
            Hu, Qinghua and Ling, Haibin},
    journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
    volume={44}, number={11}, pages={7380--7399}, year={2021}
}