🌿 Robust Tracking against Adversarial Attacks
Shuai Jia, Chao Ma, Yibing Song and Xiaokang Yang
European Conference on Computer Vision (ECCV), 2020
Deep convolutional neural networks (CNNs) are vulnerable to adversarial attacks.
- We propose to generate adversarial examples to deteriorate the performance for visual object tracking.
- Conversely, we propose to defend deep trackers against adversarial attacks that eliminate their effect to alleviate performance drops caused by the adversarial attack.
- We choose two typical trackers, DaSiamRPN and RT-MDnet.
The environment follows the tracker you intend to attack:
-
The specific setting and pretrained model for DaSiamPRN can refer to Code_DaSiamRPN
-
The specific setting and pretrained model for RT-MDNet can refer to Code_RT-MDNet
OTB2015 OP / DP |
VOT2018 A / R / EAO |
VOT2016 A / R / EAO |
UAV123 OP / DP |
|
---|---|---|---|---|
DaSiamRPN | 0.658 / 0.886 | 0.585 / 0.272 / 0.380 | 0.622 / 0.214 / 0.418 | 0.592 / 0.791 |
DaSiamRPN+RandAtt | 0.586 / 0.799 | 0.571 / 0.529 / 0.223 | 0.606 / 0.303 / 0.336 | 0.572 / 0.769 |
DaSiamRPN+Att | 0.050 / 0.050 | 0.536 / 1.447 / 0.097 | 0.521 / 1.631 / 0.078 | 0.026 / 0.045 |
DaSiamRPN+Att+Def | 0.473 / 0.639 | 0.579 / 0.674 / 0.195 | 0.581 / 0.722 / 0.211 | 0.465 / 0.639 |
DaSiamRPN+Def | 0.658 / 0.886 | 0.584 / 0.253 / 0.384 | 0.625 / 0.224 / 0.439 | 0.592 / 0.792 |
OTB2015 OP / DP |
VOT2018 A / R / EAO |
VOT2016 A / R / EAO |
UAV123 OP / DP |
|
---|---|---|---|---|
RT-MDNet | 0.643 / 0.876 | 0.533 / 0.567 / 0.176 | 0.567 / 0.196 / 0.370 | 0.512 / 0.754 |
RT-MDNet+RandAtt | 0.559 / 0.753 | 0.503 / 0.871 / 0.137 | 0.550 / 0.452 / 0.235 | 0.491 / 0.728 |
RT-MDNet+Att | 0.131 / 0.140 | 0.475 / 1.611 / 0.076 | 0.469 / 0.928 / 0.128 | 0.079 / 0.128 |
RT-MDNet+Att+Def | 0.420 / 0.589 | 0.515 / 1.021 / 0.110 | 0.531 / 0.494 / 0.225 | 0.419 / 0.620 |
RT-MDNet+Def | 0.644 / 0.883 | 0.529 / 0.538 / 0.179 | 0.540 / 0.168 / 0.364 | 0.513 / 0.757 |
🌿 All raw results are available. [Google_drive] [Baidu_Disk] Code: 5ex9
🌿 The code of adversarial attack on DaSiamRPN is released!!
- You should download the OTB2015 dataset in
data
folder. - Please download the pretrained model in Code_DaSiamRPN.
Test the original performance on OTB2015 dataset, please using the follwing command.
cd DaSiamRPN/code
python test_otb.py
Test the adversarial attack performance on OTB2015 dataset, please using the follwing command.
cd DaSiamRPN/code
python test_otb_attack.py
Test the adversarial defense performance on OTB2015 dataset, please using the follwing command.
cd DaSiamRPN/code
python test_otb_defense.py
-v
can be used to visualize the tracking results.
If any part of our paper and code is helpful to your work, please generously citing:
@inproceedings{jia-eccv20-RTAA,
title={Robust Tracking against Adversarial Attacks},
author={Jia, Shuai and Ma, Chao and Song, Yibing and Yang, Xiaokang},
booktitle={European Conference on Computer Vision},
year={2020}
}
@inproceedings{zhu-eccv18-dasiamrpn,
title={Distractor-aware Siamese Networks for Visual Object Tracking},
author={Zhu, Zheng and Wang, Qiang and Li, Bo and Wu, Wei and Yan, Junjie and Hu, Weiming},
booktitle={European Conference on Computer Vision},
year={2018}
}
@InProceedings{jung-eccv19-rtmdnet,
author = {Jung, Ilchae and Son, Jeany and Baek, Mooyeol and Han, Bohyung},
title = {Real-Time MDNet},
booktitle = {European Conference on Computer Vision (ECCV)},
month = {Sept},
year = {2018}
}
Thank you!
Licensed under an MIT license.