Towards Universal Physical Attacks on Single Object Tracking

Recent studies show that small perturbations in video frames could misguide single object trackers. However, such attacks have been mainly designed for digital-domain videos (i.e., perturbation on full images), which makes them practically infeasible to evaluate the adversarial vulnerability of trackers real-world scenarios. Here we made first step towards physically feasible against visual tracking real scenes with a universal patch camouflage Fundamentally different from physical detection, essence lies feature matching between search image and templates, therefore specially design maximum textural discrepancy (MTD), resolution-invariant target location-independent de-matching loss. The MTD distills global information template images at hierarchical scales prior performing attacks. Moreover, two shape attacks, regression dilation shrinking, generate stronger more controllable Further, employ set transformations simulate diverse wild. Experimental results effectiveness SiamMask SiamRPN++ both digital scenes.