Siamese Visual Tracking with Feature Enhancement Fusion
Zhixi Wu, Baichen Liu, Shunzhi Zhu · 2024
In object tracking tasks, the use of a Siamese-based approach to construct trackers inevitably involves a crucial step - cross-correlation operations, which are employed to assess the similarity relationship between the template and the search area. However, this method is limited to obtaining only local information, lacking in global contextual understanding. Inspired by the Transformer’s capability to capture long-term dependencies, we propose a novel Siamese visual target tracking method with an attention mechanism, named SiamTr. Specifically, we design a feature enhancement fusion network to replace the cross-correlation computation, composed of a Hybrid Attention Module and a Cross-Attention Fusion Module. Additionally, we incorporate an Attention Feature Fusion(AFF) to merge shallow and deep features, thus acquiring diverse information. Experiments demonstrate that our proposed method achieves competitive results in multiple benchmark datasets, including OTB100, UAV123, and LaSOT.