A Asymmetric Attention Siamese Network for Visual Object Tracking
Shaozhe Guo, Yong Li, Youshan Zhang, Qiming Liang, Yang Kaikai, Zhai Kewen · 2021
To make the visual tracking aims algorithm more adaptive to shallow feature extraction, this paper proposes a Siamese network based trackers that can enhance the robustness of target trackers. Combined with attention model and feature fusion, a new asymmetric convolution structure is applied, which can effectively improve the performance of Siamese networks. Ablation experiments have proved that this model behaves well while features are of different depths. Finally, experiments are carried out on the OTB2015 benchmarks set, which show that performance of this algorithm exceeds other benchmark algorithms. Compared with the previous SiamFC, the success plot of this increased by 3.4%, and real-time tracking speed is also achieved. The advantages of the proposed algorithm are verified.