Robust Siamese Tracking with Asymmetrical Feature Processing Network

Abstract

The Siamese tracker consists of two components: a classification and a regression networks. Despite their different roles, most Siamese trackers have similar feature fusion modules in the two networks, leading to the neglect of their unique characteristics. In this work, we experimentally discover that the two networks place different levels of emphasis on different types of information. Specifically, regression tends to rely on semantic information, while classification places more emphasis on global information. Therefore, we propose a new tracking structure named SGTrack, which includes a semantic augmentation fusion (SAF) for regression and a global relevance fusion (GRF) for classification. It allows us to unlock the full potential of both networks. The experimental results of our method on five benchmarks provide evidence of a notable improvement in tracking performance, while preserving real-time speed.