SiamFFN: Siamese Feature Fusion Network for Visual Tracking

Siamese network-based trackers have developed rapidly in the field of visual object tracking recently. Many currently use rely on result fusion to combine classification map and regression map. However, these maps are obtained from multi-level feature independent each other. It is inappropriate flawed fusion. Additionally, module other, which leads misalignment. In this paper, we propose a feature-fusion approach that involves fusing similarity response using novel scale attention mechanism subsequently decoding features. To reduce misalignment produce more precise results, suggest Classification Supervised Regression Loss (CSRL), train model. Experiments conducted three challenging benchmark datasets show method outperforms current models terms both performance efficiency, running at 40 fps.