Siamese anchor-free network based on dual attention mechanism for visual tracking

Abstract

In order to improve the ability of extracting discriminant features and add valid location information, a new Siamese anchor-free network based on dual attention mechanism is proposed—DASN. The method introduces the coordinate channel attention module and the spatial attention module to get the context features that contain the precise location information of the target. DASN can enhance the feature extraction by modeling the dependency between channels and positions, thus achieve the improvement of the accuracy of classification and locating and the robustness of tracking. Experimental results show that the proposed method achieves the performance improvement on datasets VOT2018, OTB-2013 and OTB-2015, and also meets the real-time requirements.