SiamSGA: Siamese Symmetric Graph Attention Tracking

Visual tracking is commonly approached through similarity estimation between a template and a search region in recent Siamese-based trackers. These trackers employ cross-correlation to generate similarity maps from pairs of feature maps, achieving commendable performance in visual tracking. Despite their success, these cross-correlation methods exhibit certain limitations. The presence of redundant background information can distract trackers from the target, while scale mismatches between the template and the candidate can lead to an overemphasis on global features. In this paper, we introduce a novel approach for visual tracking: the Symmetric Graph Attention Network (SiamSGA). SiamSGA is designed to effectively capture both global and local information. Our approach establishes part-to-part and integral-to-integral connections between feature maps, facilitating the encoding of more valuable information from two distinct branches. Extensive experiments have been conducted on five widely recognized benchmarks, including LaSOT, UAV123, NFS30, OTB100, and NFS240. The experimental results demonstrate that our proposed tracker, SiamSGA, consistently outperforms many state-of-the-art trackers in terms of tracking accuracy.