shijli/aitod-v2

AI-TOD-v2

AI-TOD-v2, the tiny-object detection benchmark in aerial images, packed once with the official v2 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/aitod-v2")  # 11214 train / 2804 validation / 14018 test

AI-TOD cuts 28036 images of 800 x 800 pixels from xView, DOTA-v1.5, VisDrone2018-Det, Airbus Ship Detection and DIOR, and annotates eight classes whose mean object size is about 12.8 pixels. AI-TOD-v2 keeps the images and re-annotates the boxes, fixing the missed and imprecise labels of the first release (752745 instances against 700621). Only the v2 annotations are packed here.

Splits

split	images	boxes	official file
train	11214	301534	aitodv2_train.json
validation	2804	75091	aitodv2_trainval.json minus aitodv2_train.json
test	14018	376121	aitodv2_test.json

The v2 release ships train, trainval and test files and no separate validation file. The validation split is therefore trainval minus train, matched on file name. It is exactly the 2804 images of the original AI-TOD validation set, and the boxes of the train images inside trainval are identical to those of the train file, so the subtraction loses and invents nothing.

id is the AI-TOD file stem (e.g. 0000001_03499_d_0000006__1120_280, a VisDrone frame; xView-derived ids look like 1__800_0), so rows join trivially to a local AI-TOD installation, where the file is <id>.png.

Columns

column	type	meaning
id	string	the AI-TOD file stem
image	image	the PNG, bytes unchanged
width, height	int32	image size in pixels (800 x 800, a few images a pixel or two shorter)
objects.category	ClassLabel(8)	the official v2 category id, see below
objects.bbox	4 x float32	the official [x, y, w, h], absolute pixels, top-left origin

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"]):
    ...

The annotations are the official ones, verbatim

Each split holds exactly as many boxes as the v2 json assigns to its images, and every box is the json's own [x, y, w, h], floats included. Nothing is filtered: the 51 boxes with a zero width or height (40 train, 9 validation, 2 test) and the 28 images without any box (10 train, 18 test) are kept, so that which boxes to train on is a decision left to the consumer.

Three json fields are not packed because they carry no information: area is w * h of the box to floating point precision on every annotation, segmentation is the box's own four corners, and iscrowd is 0 everywhere. All three were checked over every annotation of every split before being dropped.

The category ClassLabel is indexed by the official AI-TOD-v2 id:

id	name	id	name
0	airplane	4	swimming-pool
1	bridge	5	vehicle
2	storage-tank	6	person
3	ship	7	wind-mill

The first AI-TOD release numbered the same eight classes 1 to 8 in the same order; v2 numbers them 0 to 7, and so does this dataset.

How this was packed

The repository includes create_dataset.py, which produced these parquet files from an AI-TOD folder holding the generated images and the three v2 json files. Images are copied through byte for byte. Before pushing, the build was checked against the source: every split holds exactly as many boxes as the json assigns to its images, every recorded size matches the PNG, and sampled records decode back to the same categories and boxes as the json.

License and attribution

AI-TOD is released by the CAPTAIN group at Wuhan University under CC BY-NC-SA 4.0, free for academic and individual research and restricted for commercial use. Its images come from xView, DOTA-v1.5, VisDrone2018-Det, Airbus Ship Detection and DIOR, each with its own non-commercial research terms. This packaging adds no rights: use it under those same terms and cite the original work.

@article{xu2022detecting,
    title={Detecting Tiny Objects in Aerial Images: A Normalized Wasserstein Distance and A New Benchmark},
    author={Xu, Chang and Wang, Jinwang and Yang, Wen and Yu, Huai and Yu, Lei and Xia, Gui-Song},
    journal={ISPRS Journal of Photogrammetry and Remote Sensing},
    volume={190}, pages={79--93}, year={2022}
}
@inproceedings{wang2021tiny,
    title={Tiny Object Detection in Aerial Images},
    author={Wang, Jinwang and Yang, Wen and Guo, Haowen and Zhang, Ruixiang and Xia, Gui-Song},
    booktitle={International Conference on Pattern Recognition (ICPR)},
    pages={3791--3798}, year={2021}
}