/ABA

We propose the adversarial blur attack (ABA) against visual object tracking.

Primary LanguagePythonMIT LicenseMIT

ABA

We propose the adversarial blur attack (ABA) against visual object tracking. The ICCV link: https://arxiv.org/abs/2107.12085 and, https://openaccess.thecvf.com/content/ICCV2021/papers/Guo_Learning_To_Adversarially_Blur_Visual_Object_Tracking_ICCV_2021_paper.pdf

Motion Blur Systhensis for Visual Object Tracking

JAMANet for One-step Adversarial Blur Attack

Results

Case1

Case2

Case3

Case4

Case5

Case6

Usage

Our implementation is based on PySOT, You need to follow their installation steps for environment setup.

You can enter the target tracker's directory in the experiments folder, run the attack/evaluate experiment with the following command:

# attack
CUDA_VISIBLE_DEVICES=0 python -u ../../tools/attack_test.py --snapshot  model.pth  --dataset OTB100  --config config.yaml
CUDA_VISIBLE_DEVICES=0 python -u ../../tools/attack_dimp.py --snapshot  model.pth  --dataset OTB100  --config config.yaml
CUDA_VISIBLE_DEVICES=0 python -u ../../tools/attack_kys.py --snapshot  model.pth  --dataset OTB100  --config config.yaml

# evaluate
python ../../tools/eval.py  --tracker_path ./results   --dataset OTB100   --num 10 --tracker_prefix 'cl'
python ../../tools/eval.py  --tracker_path ./results   --dataset VOT2018  --num 10 --tracker_prefix 'clean'
python ../../tools/eval.py  --tracker_path ./results   --dataset UAV      --num 10 --tracker_prefix 'guo'
python ../../tools/eval.py  --tracker_path ./results   --dataset VOT2019  --num 10--tracker_prefix 'csa'

Bibtex

@inproceedings{guo2021learning,
      title={Learning to Adversarially Blur Visual Object Tracking},
      author={Qing Guo and Ziyi Cheng and Felix Juefei-Xu and Lei Ma and Xiaofei Xie and Yang Liu and Jianjun Zhao},
      year={2021},
      booktitle={ICCV}
}