Response Feedback for Task Semantic Space Consistent Siamese Tracking

Siamese trackers typically consist of a classification branch for separating the target and the background, and a regression branch for predicting the target's bounding box. However, sometimes these two branches produce conflicting results, such as accurate classification but imprecise bounding box prediction, or vice versa, which do reduce the tracker's performance. To address this issue, a remodulated Siamese tracker that maintains consistency between the classification and regression branches in the task-semantic space is proposed, abbreviated as task semantic space consistent Siamese tracker and named TSSC-Siam. First, consistency semantic integration (CSI) framework is designed to maintain result consistency between the classification and regression branches. A Response Activation Attention (RAA) module is introduced in CSI, which learns to activate a relevant class map for each task. By fusing the two class maps, a shared spatial region that forces the two branches to focus on the same area is created, thus mitigating conflicts between the two tasks. Experiments on OTB100, GOT10K, UAV123, and LaSOT benchmarks demonstrate that the approach achieves comparable performance to some popular methods.