Accumulation of adversarial examples for underwater visual object tracking

Abstract

Recent studies show that small perturbations in video frames could misguide the deep learning-based visual object trackers. In this paper, we first attempt to generate an accumulation of adversarial examples for underwater VOT (Visual object tracking). This is the first attempt at underwater VOT attack. The data used are the URPC (Underwater Robot Picking Contest) in 2017 and 2018, and the fish2 subset of the VOT2019(Visual Object Tracking Challenge in 2019). We generate adversarial examples by minimizing the L2 total loss function which we designed. Experiments show that this attack method can achieve an effective attack, which can reduce the success rate of at least 48.6% of the DaSiamRPN(Distractor-aware SiamRPN) trackers.