GhostSiamRPN++: Optimizing SiamRPN++ via GhostModule

Abstract

SiamRPN++ is a Siamese network based tracker in object tracking and it has excellent performance. However, SiamRPN++ is a ResNet-driven tracker with huge amounts of parameters and calculations. It has extremely strict requirements for hardware and devices without professional GPU are difficult to meet the requirements. So SiamRPN++ should become more lightweight by optimization. In this paper, we propose a lightweight tracker named GhostSiamRPN++ by optimizing SiamRPN++ and replacing the ResNet backbone with a lightweight backbone build by GhostModule. When compared to SiamRPN++, GhostSiamRPN++ reduce 84% parameters, 92% calculations and 52% memory usage and it boost the FPS by 55% on GPU and by 545% on CPU. Large scale optimization does not bring great loss of performance. When tested on VOT and OTB datasets, GhostSiamRPN++ shows excellent performance in all videos and leading performance in some specific videos and it has 4.5% to 13.6% lower accuracy than SiamRPN++.