IoUattack

🌿 IoU Attack: Towards Temporally Coherent Black-Box Adversarial Attack for Visual Object Tracking

Shuai Jia, Yibing Song, Chao Ma and Xiaokang Yang

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021.

Introduction

We observe that the increase of noise level positively correlates to the decrease of IoU scores, but their directions are not exactly the same.

• Our IoU attack seeks to inject the lowest amount of noisy perturbations at the same contour line of IoU score for each iteration.
• We choose three representative trackers with different structures, SiamRPN++, DiMP and LTMU, respectively.

Results

Result for SiamRPN++ on multiple datasets

	VOT2019 A / R / EAO	VOT2018 A / R / EAO	VOT2016 A / R / EAO	VOT2018lt F-score	OTB2015 OP / DP	NFS30 OP / DP
SiamRPN++	0.596 / 0.472 / 0.287	0.602 / 0.239 / 0.413	0.643 / 0.200 / 0.461	0.625	0.695 / 0.905	0.509 / 0.601
SiamRPN++(Random)	0.591 / 0.727 / 0.220	0.587 / 0.365 / 0.301	0.632 / 0.340 / 0.331	0.553	0.631 / 0.818	0.466 / 0.550
SiamRPN++(Attack)	0.575 / 1.575 / 0.124	0.568 / 1.171 / 0.129	0.605 / 0.802 / 0.183	0.453	0.499 / 0.644	0.394 / 0.446

Result for DiMP on multiple datasets

	VOT2019 A / R / EAO	VOT2018 A / R / EAO	VOT2016 A / R / EAO	VOT2018lt F-score	OTB2015 OP / DP	NFS30 OP / DP
DiMP	0.568 / 0.277 / 0.332	0.574 / 0.145 / 0.427	0.599 / 0.140 / 0.449	0.609	0.671 / 0.869	0.614 / 0.729
DiMP(Random)	0.567 / 0.373 / 0.284	0.560 / 0.202 / 0.363	0.592 / 0.168 / 0.404	0.555	0.659 / 0.860	0.591 / 0.710
DiMP(Attack)	0.474 / 0.641 / 0.195	0.507 / 0.400 / 0.248	0.536 / 0.374 / 0.256	0.443	0.592 / 0.791	0.545 / 0.658

Result for LTMU on multiple datasets

	VOT2019 A / R / EAO	VOT2018 A / R / EAO	VOT2016 A / R / EAO	VOT2018ltT F-score	OTB2015 OP / DP	NFS30 OP / DP
LTMU	0.625 / 0.913 / 0.201	0.624 / 0.702 / 0.195	0.661 / 0.522 / 0.236	0.691	0.672 / 0.872	0.631 / 0.764
LTMU(Random)	0.623 / 1.073 / 0.175	0.622 / 0.805 / 0.178	0.646 / 0.592 / 0.233	0.657	0.622 / 0.815	0.579 / 0.699
LTMU(Attack)	0.576 / 1.470 / 0.150	0.590 / 1.320 / 0.120	0.604 / 0.904 / 0.170	0.589	0.517 / 0.712	0.462 / 0.559

🌿 All raw results are available. [Google_drive]

Code

🌿 The code of IoU attack for SiamRPN++ is released!!

• You should put the datasets into pysot/testing_dataset folder.
• Please download the pretrained model and set the environments of SiamPRN++.
• See SiamRPN++ for more details.

Test the original performance on VOT2018 dataset, please use the following command.

cd pysot/experiments/siamrpn_r50_l234_dwxcorr
python -u ../../tools/test_original.py 	\
	--snapshot model.pth 	\ # model path
	--dataset VOT2018 	\ # dataset name
	--config config.yaml	  # config file

Test IoU attack on VOT2018 dataset, please use the following command.

cd pysot/experiments/siamrpn_r50_l234_dwxcorr
python -u ../../tools/test_IoU_attack.py 	\
	--snapshot model.pth 	\ # model path
	--dataset VOT2018 	\ # dataset name
	--config config.yaml	  # config file

For the adversarial attack of other datasets, you should change the dataset name as mentioned above.

Demo

       

Citation

If any part of our paper and code is helpful to your work, please generously citing:

@inproceedings{jia-cvpr21-iouattack,
  title={IoU Attack: Towards Temporally Coherent Black-Box Adversarial Attack for Visual Object Tracking},
  author={Jia, Shuai and Song, Yibing and Ma, Chao and Yang, Xiaokang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2021}
}

Thank you :)

Reference

We choose three representative trackers, SiamRPN++, DiMP and LTMU. The original code of these trackers are list as follows:

• SiamRPN++: https://github.com/STVIR/pysot
• DiMP: https://github.com/visionml/pytracking
• LTMU: https://github.com/Daikenan/LTMU

We also refer to the code of Boundary Attack for IoU attack.

• Boundary Attack: https://github.com/greentfrapp/boundary-attack

Thanks for their wonderful works!

License

Licensed under an MIT license.