An Anchor-Free Siamese Network with Multi-Template Update for Object Tracking

Siamese trackers are widely used in various fields for their advantages of balancing speed and accuracy. Compared with the anchor-based method, anchor-free-based approach can reach faster speeds without any drop precision. Inspired by network anchor-free idea, an (AFSN) multi-template updates object tracking is proposed. To improve performance, a dual-fusion method adopted which multi-layer features multiple prediction results combined respectively. The low-level feature maps concatenated high-level to make full use both spatial semantic information. as stable possible, final obtained combining results. Aiming at template update, high-confidence update mechanism used. average peak correlation energy determine whether should be updated. We implement per-pixel manner, computes category bounding boxes directly. Experimental indicate that overlap success rate proposed algorithm increase about 5% 10%, respectively, compared SiamRPN++ when running on dataset GOT-10k (Generic Object Tracking Benchmark).