AI-TOD: Tiny Object Detection in Aerial Images

AI-TOD is a benchmark dataset for tiny object detection in aerial images.

• Paper: Tiny Object Detection in Aerial Images (ICPR 2021)
• Source repository: https://github.com/jwwangchn/AI-TOD
• Images: 28,036 aerial images
• Instances: 700,621 annotated instances
• Categories: 8
• Average object size: 12.8 pixels

Splits

Split	Images
train	11,214
val	2,804
test	14,018
Total (trainval = train + val = 14,018; plus test)	28,036

Annotation counts verified against the paper (v1, full annotation set): trainval 14,018 images / 353,004 instances, test 14,018 images / 347,617 instances — 28,036 images / 700,621 instances in total.

Files

File	Size (bytes)	Contents
train.zip.part1 … train.zip.part4	3,221,225,472 ×3 + 163,818,951	Byte-split parts of train.zip (9,827,495,367 bytes in total). Training images (11,214 PNG images, plus per-image label .txt files) under train/images/
val.zip	2,461,815,071	Validation images (2,804 PNG images, plus per-image label .txt files) under val/images/
test.zip.part1 … test.zip.part4	3,221,225,472 ×3 + 2,649,652,405	Byte-split parts of test.zip (12,313,328,821 bytes in total). Test images (14,018 PNG images) under test/images/
annotations_v1.zip	36,859,877	v1 COCO-style annotations: annotations/aitod_train_v1.json, annotations/aitod_val_v1.json, annotations/aitod_trainval_v1.json, annotations/aitod_test_v1.json
annotations_v2.zip	11,812,370	v2 annotations: annotations/aitod_train_v2.json, annotations/aitod_val_v2.json

The archives keep the official original file names and internal structure.

Reassembling the split archives: train.zip and test.zip are uploaded as byte-split parts because of their size. Concatenate the parts to restore the original archive byte-for-byte, then unzip as usual:

cat train.zip.part1 train.zip.part2 train.zip.part3 train.zip.part4 > train.zip
cat test.zip.part1 test.zip.part2 test.zip.part3 test.zip.part4 > test.zip

SHA-256 of the reassembled archives:

• train.zip: c399cc32c39eba5fd0dacdcde1ec12408f862a52c20aa295b87faeb01fd9bb4d
• test.zip: f33fec10e24b07d13273299a434ffef30ca3ad20e39c2931466dc415b9d529ec

Annotation versions: v1 vs v2

• v1 is the original annotation set released with the paper, covering train, val, trainval and test.
• v2 is the revised annotation set from AI-TOD-v2, which re-annotates / refines the labels for the train and val splits (this archive contains aitod_train_v2.json and aitod_val_v2.json). v1 and v2 share the same images. See https://chasel-tsui.github.io/AI-TOD-v2/ for details of the v2 annotations.

License

The dataset is released under CC BY-NC-SA 4.0 as declared by the AI-TOD authors (non-commercial use, share-alike). The accompanying code in the source repository is under the MIT license — the dataset license and the code license are separate.

Citation

If you use this dataset, please cite the original paper: Tiny Object Detection in Aerial Images, ICPR 2021, and refer to the source repository https://github.com/jwwangchn/AI-TOD.

Related Publications

Tiny object detection research by the AI-TOD authors:

1. Jinwang Wang, Wen Yang, Haichao Li, Haihui Ye, Lei Yu. Tiny Object Detection in Aerial Images. 25th International Conference on Pattern Recognition (ICPR), 2021. (The AI-TOD dataset paper)
2. Jinwang Wang, Chang Xu, Wen Yang, Lei Yu. A Normalized Gaussian Wasserstein Distance for Tiny Object Detection. arXiv:2110.13389, 2021. [arXiv]
3. Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, Gui-Song Xia. Detecting Tiny Objects in Aerial Images: A Normalized Wasserstein Distance and a New Benchmark. ISPRS Journal of Photogrammetry and Remote Sensing, vol. 190, pp. 79–93, 2022. [arXiv] (Introduces AI-TOD-v2)
4. Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, Gui-Song Xia. RFLA: Gaussian Receptive Field based Label Assignment for Tiny Object Detection. European Conference on Computer Vision (ECCV), pp. 526–543, 2022. [arXiv]
5. Chang Xu, Jian Ding, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, Gui-Song Xia. Dynamic Coarse-To-Fine Learning for Oriented Tiny Object Detection. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7318–7328, 2023. [arXiv]

@inproceedings{wang2021aitod,
  title     = {Tiny Object Detection in Aerial Images},
  author    = {Wang, Jinwang and Yang, Wen and Li, Haichao and Ye, Haihui and Yu, Lei},
  booktitle = {2020 25th International Conference on Pattern Recognition (ICPR)},
  year      = {2021}
}
@article{wang2021nwd,
  title   = {A Normalized Gaussian Wasserstein Distance for Tiny Object Detection},
  author  = {Wang, Jinwang and Xu, Chang and Yang, Wen and Yu, Lei},
  journal = {arXiv preprint arXiv:2110.13389},
  year    = {2021}
}
@article{xu2022benchmark,
  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{xu2022rfla,
  title     = {RFLA: Gaussian Receptive Field based Label Assignment for Tiny Object Detection},
  author    = {Xu, Chang and Wang, Jinwang and Yang, Wen and Yu, Huai and Yu, Lei and Xia, Gui-Song},
  booktitle = {European Conference on Computer Vision (ECCV)},
  pages     = {526--543},
  year      = {2022}
}
@inproceedings{xu2023dcfl,
  title     = {Dynamic Coarse-To-Fine Learning for Oriented Tiny Object Detection},
  author    = {Xu, Chang and Ding, Jian and Wang, Jinwang and Yang, Wen and Yu, Huai and Yu, Lei and Xia, Gui-Song},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {7318--7328},
  year      = {2023}
}