Adversarial Attacks on Object Detection in Optical Remote Sensing Images: Benchmarks

To facilitate the evaluation of the following research regarding adversarial attacks on object detection in O-RSIs, we sample 2000 images from the testing subset of DIOR and the validation subset of DOTA, respectively, dubbed DIOR-A and DOTA-A. The class-wise instance distributions of them are exhibited below. Here, both the class-wise instance distribution histograms and their corresponding Kernel Density Estimation (KDE) curves are plotted. These datasets are firstly leveraged in Threatening Patch Attacks on Object Detection in Optical Remote Sensing Images for the evaluations. For the training details of the victim detectors, we refer our users TPA for more detailed descriptions.

✨ News

[Apr 18, 2024]

• Codes of Threatening Patch Attacks on Object Detection in Optical Remote Sensing Images has been updated at plpl2019/TPA. Thanks very much for the selfless dedication of Lei Pei to this work! Meanwhile, I sincerely thank all the colleagues who have shown their interest in this project. Thank you for your support and understanding!

Architecture

These datasets can be accessed at BaiduNetDisk and Google Drive, where the architecture of this file is:

.datasets
├── README.md
├── dist.png
├── DIOR-A
      ├──images
      └──labelTxt
└── DOTA-A
      ├──images
      └──labelTxt

Here, the image file and the coresponding label file are named with the same series number, e.g.,

An image file with the file name of ./datasets/DIOR-A/images/11726.png,

then its label file is ./datasets/DIOR-A/labelTxt/11726.txt

Acknowledgement

We sincerely thank the authors of DIOR and DOTA for their great contributions to optical remote sensing images understanding. Besides, DIOR-A and DOTA-A are mainly built by Lei Pei and Xuxiang Sun, with the guidance of Prof. Gong Cheng.

Citation

If you use these datasets, please cite the following:

@article{sun2023threatening,
  title={Threatening Patch Attacks on Object Detection in Optical Remote Sensing Images},
  author={Sun, Xuxiang and Cheng, Gong and Pei, Lei and Li, Hongda and Han, Junwei},
  journal={IEEE Trans. Geosci. Remote Sens.},
  volume={61},
  pages={1-10},
  year={2023}
}

Contact

If you can not access the download links, feel free to contact xuxiangsun@mail.nwpu.edu.cn (English and Chinese are all ok).