Memory Mechanisms for Discriminative Visual Tracking Algorithms With Deep Neural Networks

Abstract

Deep-neural-networks-based online visual tracking methods have achieved state-of-the-art results. One of the core components of these methods is the memory pool, in which a number of samples consisting of image patches and the corresponding labels are stored to update the online tracking network. Hence, the mechanism of updating the stored samples determines the performance of the tracking method. In this paper, a novel memory mechanism is proposed to control the writing and reading accesses of the memory pool using credit assignment network ${H}$ , which learns features of the target object. This memory mechanism comprises the writing and reading mechanisms. In the writing mechanism, network ${H}$ produces credits for the current tracked object and the samples in the memory pool. This ensures that the reliable samples are written into the memory pool and the unreliable samples are replaced if the memory pool is full. In the reading mechanism, network ${H}$ assigns an importance score to each sample selected to update the online tracking network. The state-of-the-art tracking methods with and without the proposed memory mechanism are evaluated on the CVPR2013 and OTB100 benchmarks. The experimental results demonstrated that the proposed memory mechanism improves tracking performance significantly.