Siamese Tracker with Attention Correlation and Template Update
Xiangchen Xu, Ziyang Wan, Jing Zhou, Zhangsheng Yang · 2024
Visual object tracking is an fundamental task in computer vision. Most Siamese trackers only adopt the target image given in the initial frame as the template, and the predictions made during tracking phase are not utilized further. As the target’s appearance changes intensely throughout the video, the similarity between the template of the first frame and the current target would decrease, resulting in tracking drift. To address this issue, a tracker with attention module and online template update module is proposed in this paper. The online update module meticulously discerns top-tier templates from the prognostic outcomes throughout the tracking phase. The attention module is employed to embed the template, augmenting the significance of crucial features. The test results on the OTB100 and UVA123 datasets show that our proposed tracker has achieved competitive tracking performance in the face of challenges such as interference, occlusion and deformation.