Adversarial attack can help visual tracking

Sungmin Cho, Hyeseong Kim, Ji Soo Kim, Hyomin Kim, Junseok Kwon

Abstract: We present a novel noise-injected Markov chain Monte Carlo (NMCMC) method for visual tracking, which enables fast convergence through adversarial attacks. The proposed NMCMC consists of three steps: noise-injected proposal, acceptance, and validation. We intentionally inject noise into the proposal function to cause a shift in a direction that is opposite to the moving direction of a target, which is viewed in the context of an adversarial attack. This noise injection mathematically induces the proposed visual tracker to find a target proposal distribution using a small number of samples, which allows the tracker to be robust to drifting. Experimental results demonstrate that our method achieves state-of-the-art performance, especially when severe perturbations caused by an adversarial attack exist in the target state.