Siamese Networks with Discriminant Correlation Filters and Channel Attention

Abstract

In recent years, Discriminant Correlation Filters (DCF) based methods have performed well on online object tracking. And Fully-convolutional Siamese network becomes a dominant approach to real-time object tracking. In this work, we build a two-fold Siamese network, namely SiamDCF, to learn the convolutional features and perform the correlation tracking process with channel attention simultaneously. We train these two branches of SiamDCF separately, ensuring their heterogeneous features. We treat DCF as a correlation filter layer, and the layer outputs the response map of object location. This branch learns filters which extract semantic features and perform well in situations, such as deformation and motion blur, as a complement to the original SiamFC. In particular, we introduce the channel attention module to the network. The architecture and channel attention mechanism improve the tracking performance. The network is trained on the ILSVRC15 dataset for object detection in video. The proposed architecture is end-to-end and operates at frame-rates beyond real-time. We perform comprehensive experiments on OTB2013 benchmark, and the proposed tracker achieves high performance.