Dynamic Template Update Mechanism for Visual Object Tracking
Yongqi Hu, Zhongwang Cai, ZiJun Nie, Zizheng Yu, Ying Shan Cao, Zhijian Yin, Zhen Yang · 2023
Given the requirements for robust target classification and accurate target state estimation in visual tracking, SiamFC++ proposes a set of practical guidelines for designing high-performance general-purpose trackers by considering the special nature of visual tracking problems. Inspired by dynamic modules, We propose an empirical method for integrating a dynamic module into the image input, which is concatenated with the template module after feature maps are extracted by the backbone network. Since the position and shape of the object can change significantly within a video sequence, the added dynamic module can better focus on the target region of the feature map to obtain better similarity maps. Extensive experiments and comparisons demonstrate that our simple and effective method achieves reliable results on the benchmarks of LaSOT, TrackingNet, and GOT-10K and provides a significant speed advantage in real-time.