Dual-Attention Feature Decoupling in Siamese Networks for Object Tracking

Object tracking have been made significant progress in Siamese-based trackers, however, it is still difficult to track the target under scale variations, occlusions and unknown deformations. To solve this issue, we propose to decouple feature in tracking-heads by dual-attention to alleviate the conflict between the required features of classification and regression tasks. Specifically, we introduce channel attention to enhance high-level semantic feature for classification and adopt deformable attention to stress high-resolution local detail information for regression, separately. Besides, we propose multi-stage losses strategy to fuse multi-level features for effectively optimizing the classification and regression tasks after decoupling tracking-heads to improve tracking performance. Extensive experiments on visual tracking benchmarks including VOT2019 and OTB100, demonstrate that SiamDTH achieves state-of-the-art performance with a considerable real-time speed.