Reinspecting Classification and Regression in the Sibling Head for Visual Tracking

Abstract

The sibling head has been widely used in Siamese-based trackers, however, the structure of the sibling head is the same between classification and regression tasks, which limits the ability of the tracker to obtain more robust and accurate prediction. To solve this issue, we reinspect the network structure of tracking-head for classification and regression tasks, since recognizing the target category needs translation invariant feature while the position-sensitive information facilitates target bounding-box regression task. Further, we propose a differenti-ated tracking-head network, named SiamDTH, by exploiting the feature response module (FRM) and the differentiated sibling head (DSH) to alleviate misalignments between classification and regression task domains. Extensive experiments on visual tracking benchmarks including VOT2019 and OTB100 demon-strate that SiamDTH achieves state-of-the-art performance with a considerable real-time speed. Our source code is available at: https://github.com/x10312/SiamDTH.