Learning Channel Attention in Frequency Domain for Visual Tracking

Abstract

Visual tracking based on attention mechanism, especially channel attention, has achieved great success. However, they do not make full use of the correlation between the feature map channels and use the feature information to model the target appearance. In this paper, under the Siamese network framework, we propose a simple and effective tracking algorithm based on channel attention mechanism in the frequency domain, referred to as SiamFCA. The frequency channel attention preprocesses the feature map obtained through the fully convolutional network (FCN). It divides and weights the feature maps in the frequency domain, and recalibrates the channel feature response by modeling the correlation between the feature channels. The proposed model further enhances the feature information of the extracted target image, reduces the model complexity and computational burden. The results of the experiment on the two benchmark datasets of OTB2015 and VOT2016 show that the proposed tracker based on the frequency channel attention mechanism is better than many state-of-the-art trackers.