An efficient method to fool and enhance object tracking with adversarial perturbations

Although current Siamese-based trackers have achieved impressive performance in visual object tracking by balancing accuracy and speed, their sensitivity to adversarial perturbation makes them face many risks in practical applications ( e.g., automatic drive). In this paper, we first propose a general attack method against Siamese-based trackers, and then carefully design an online augmentation strategy based on adversarial examples to improve the tracking accuracy. To ensure both attack and augmentation trackers simultaneously, two opposing losses are proposed to push and pull the gap between the template patch and search regions. Specifically, most Siamese-based trackers employ the cross-correlation module to associate features of two branches. The similarity between the template patch and the search region will be reflected on the cross-correlation map, which is the focus of our method. SiamCAR is the main research object of this paper; the tracking accuracy after deception and enhancement is reduced by 64.62% and improved by 0.88%, respectively. We also transfer our attack method to SiamRPN++, SiamBAN, SiamGAT, and TrDiMP. Extensive experiments on the popular benchmark datasets indicate the effectiveness and universality of our method.