Residual Attention SiameseRPN for Visual Tracking

Xu Cheng,Enlu Li,Zhangjie Fu

DOI: https://doi.org/10.1007/978-3-030-60639-8_34

2020-01-01

Abstract:Visual tracking demands to perform the accurate object location given the object state of the first frame. The existing methods have proposed various ways to handle the challenging problems, yet few of them take the relationship between shallow features and deep semantic features into account. Based on an extensive analysis, we first propose a residual attention SiameseRPN visual tracking method for accurate object state estimation, which introduces the correlation filter in a Siamese network framework. A novel loss function is presented to enhance the discriminative capability. Our approach is derived from three different loss terms that is capable of training a model in a few iterations. Second, we present channel attention mechanism to improve the tracking performance, which is offline trained to capture the general features in the tracking. Third, the proposed tracking model is trained in end-to-end manner and takes full advantage of both low-level representation for correlation filter and high-level semantic features for deep object representation by using multi-task learning strategy which can mine the relationship from both levels. Our approach benefits from two complementary effects. Finally, extensive evaluation and ablation studies demonstrate the effectiveness of the proposed tracking approach. Our tracker achieves state-of-the-art performance on five challenging benchmarks, which proves great potentials in balancing accuracy and speed.