Equivariant Siamese Tracking Algorithm Based on SEACT
Shijia Huang, Luping Wang, Zezheng Ye · 2021
Recently, Siamese trackers are most popular algorithms for single object tracking and have been attempted to infrared tracking. Since the dependence on the initial target template, Siamese network could not handle the scenes in the complex and changing infrared environment well, where similar interferent with various sizes makes these trackers run to drift or even tracking failure. In our previous work, we introduced the reinforcement learning model into the derivation stage of the tracker to improve the accuracy of the tracker. However, our tracker still cannot solve the problem of positional deviation led by the non-translational equivalent structure of the Siamese network when training. In this paper, we design an equivariant Siamese tracking algorithm SEACT guided by DRVGG, which is more adaptable to transformation scenarios including target scaling with the built-in scale equivariance of Siamese network. We conduct experiments on VOT-TIR2016 benchmark. The algorithm achieved excellent performance and can accurately generate zoom targets in time when the object scaling. Meanwhile, we test some sequences collected by ourselves, and there is a little improvement for some occlusion situations.