ResampleTrack: Online Resampling for Adversarially Robust Visual Tracking

Xuhong Ren, Jianlang Chen, Yue Cao, Wanli Xue, Qing Guo, Lei Ma, Jianjun Zhao, Shenyong Chen; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 8359-8363

Deep object tracking has achieved significant progress presenting impressive tracking accuracy on diverse scene videos. Nevertheless recent works have demonstrated that visual object trackers are also vulnerable to adversarial attacks designed explicitly for the tracking task. In this work we find that a naive image resampling operation could enhance the robustness of trackers under different attacks since resampling can break the adversarial patterns effectively. With this observation we propose the online resampling-based tracking defense. We perform pixel-wise resamplings for each coming frame with different shifting strategies. Moreover to fully utilize this property we propose to learn an online resampling network that can predict the pixel-wise resampling parameters (i.e. shiftings) according to different input frames automatically. We perform extensive experiments to defend against six tracking attacks against the typical tracking frameworks which demonstrates the effectiveness of our method.