Siamese Network for Visual Object Tracking with Improved Information Capture
Jun Wang, Qizhen Zhu, Kunlun Li, Sixuan Li · 2023
Siamese network based trackers could enable end-to-end tracking and show excellent performance and tracking efficiency. However, the backbone network is simple which results in insufficient ability to extract the location and semantic information. Therefor, it is difficult to handle the situation such as scale variation and deformation. Aiming the problems above, a novel Siamese based tracker named as Infc-SiamFC is proposed in this paper to efficiently obtain object location and semantic information. Firstly, channel attention is embedded in the backbone network for modeling the interdependencies among feature channels explicitly and adjusting the response value of feature channels adaptively. Secondly, a novel lightweight network named SCNet is proposed to introduce spatial information into channel attention mechanism, and improve the ability of Infc-SiamFC for extracting the location information of objects without increasing the calculation costs. Finally, the features extract from SCNet and CRNet are fused to realize feature multiplexing and further enhance the representation capability of the proposed network. The proposed Infc-SiamFC could run at 66 FPS in real-time and achieved comparable results on challenging dataset, such as OTB50, OTB100, VOT2016 and VOT2018.