Dual Attention based Siamese Network for Visual Object Tracking

Abstract

Video object tracking is a highly challenging problem, in which the initialization of the target object is given by the bounding box of first frame. The trackers based on deep Siamese network have achieved promising performance, while the robustness is still the key factor that affects the tracker's whole performance such as EAO in VOT datasets. In order to enhance the discriminability and robustness of the tracker, we introduce a dual-attentional Siamese network based tracker. In addition, we analyze the scenario that a target moves in a large scale and proposes an effective way to address this limitation. We perform extensive experiments on four public datasets. The experimental results illustrate that our novel tracker achives competitive tracking performance.