SiamRPN_IoU_attack

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.

• The IoU attack seeks to inject the lowest amount of noisy perturbations at the same contour line of IoU score for each iteration.
• We chose SiamRPN++ as the representative tracker.
• To enhance the original IoU attack we introduced dynamic Lambda adjustment which adaptively balances the spatial and temporal components of the IoU during the attack.
• To generate perturbation we used gradient based optimization, instead solely relying on random sampling.

Results

Result for IoU Attack on SiamRPN++ on VOT2018 Dataset with multiple model iterations

	Original Model	0.01 Weight Decay	0.001 Weight Decay	32 Anchors	96 Anchors	Color OFF	Blur ON
Accuracy	0.661	0.336	0.278	0.306	0.308	0.303	0.285
Robustness	1.370	1.182	1.397	1.290	1.424	1.316	2.892
EAO	0.131	0.103	0.079	0.090	0.087	0.077	0.037

Result for Enhanced IoU Attack on SiamRPN++ on VOT2018

	Original IoU	Enhanced IoU
Total Frames Lost	75	78
Average FPS	6.1	15.1
Average Time (s)	126.2	22.8

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.

Test Enhanced IoU attack

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

Thank you :)

Reference

We choose SiamRPN++ for the object tracking objective. The original code of this tracker is listed as follows:

• SiamRPN++: https://github.com/STVIR/pysot

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!