Lightweight Siamese Network for Visual Tracking via FasterNet and Feature Adaptive Fusion
Da Li, Han Tao, Huiting Zhou, Haoxiang Chai · 2024
Visual target tracking on the strength of Siamese network have made encouraging progress over recent years, providing significant advantages in the aspect of real-time and accuracy performance. Nevertheless, most of these trackers are ResN et-driver trackers with a host of parameters and high requirements for hardware. This leads to be inapplicable in some certain scenarios. In this paper, we present a lightweight but accurate tracker named SiamFaster. Using lightweight FasterNet constructed by partial convolution instead of ResN et backbone. The feature adaptive fusion module is added as a strategy that can better integration of multi-layers features and achieve better tracking performance. The new tracker is evaluated on the challenging benchmarks of OTB100 and GOT-10K and compare with some high-performance tracker. The results show that our tracker performs relatively well, with algorithm speeds up to an impressive 185 FPS.