[Paper] AI-TOD is a dataset for tiny object detection in aerial images.

[Dataset] Please download the xView trainig set and AI-TOD_wo_xview to construct the complete AI-TOD dataset!

AI-TOD comes with 700,621 object instances for eight categories across 28,036 aerial images. Compared to existing object detection datasets in aerial images, the mean size of objects in AI-TOD is about 12.8 pixels, which is much smaller than others.

You need to download the following two parts (xView training set, AI-TOD without xview) and use our end-to-end synthesis tool to generate the complete AI-TOD dataset.

• xView training set. [Website]
• AI-TOD_wo_xview. [BaiduDrive:w2dy] [Google Drive]
• E2E aitodtoolkit. [Folder]

Step 1: Download the xView training set, AI-TOD without xview, and clone the aitodtoolkit.

git clone

Step 2: Organize the downloaded files in the following way.

├─aitod
│  ├─annotations ## put the downloaded annotations of AI-TOD_wo_xview (.json)**
│  └─images ## unzip the downloaded AI-TOD_wo_xview image sets, put them (.png) in the corresponding folder
│      ├─test ## directly put the images in it without extra folder
│      ├─train 
│      ├─trainval 
│      └─val 
├─aitod_xview ## here are six files (.txt)
├─xview
│  ├─ori
│  │   └─train_images ## unzip the downloaded xView training set images, put them (.tif) here
│  └─xView_train.geojson ## the annotation file of xView training set
└─generate_aitod.py ## end-to-end tool

Step 3: Install required packages.

• Required environment

1. Python 3.7
2. mmcv

• Install wwtool

git clone https://github.com/jwwangchn/wwtool.git
cd wwtool
python setup.py develop

• Install other required packages

cd ..
cd aitodtoolkit
pip install -r requirements.txt

Step 4: Run the E2E aitodtoolkit and get AI-TOD, it might take around an hour, then the full image sets and annotations of AI-TOD can be found in the aitod folder.

python generate_aitod.py

Training, Validation and Testing sets are both publicly available now. We report the COCO style performance in the original paper, you can use the cocoapi-aitod to evaluate the model performance.

If you use this dataset in your research, please consider citing these papers.

@inproceedings{AI-TOD_2020_ICPR,
    title={Tiny Object Detection in Aerial Images},
    author={Wang, Jinwang and Yang, Wen and Guo, Haowen and Zhang, Ruixiang and Xia, Gui-Song},
    booktitle=ICPR,
    pages={3791--3798},
    year={2021},
}

@article{NWD_2021_arXiv,
  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}
}

xView Dataset

The AI-TOD dataset is licensed under the Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). Thus AI-TOD dataset are freely available for academic purpose or individual reserach, but restricted for commercial use. Besides, the underlying codes are licensed under the MIT license.