Multi-granularity Feature Fusion for Transformer-Based Single Object Tracking

The recently developed transformer has been largely explored in the research field of computer vision and especially improve the performance of single object tracking. However, the majority of current efforts concentrate on combining and enhancing convolutional neural network (CNN)-generated features and cannot fully excavating the potential of transformer. Motivated by this, we introduce multi-granularity theory into the pure transformer-based single object tracker and design a multi-granularity feature fusion module. With a view to fuse the feature of different granularity and enhance the feature representation, we design the double-branch transformer feature extractor and utilize cross-attention mechanism to fuse the feature. In our extensive experiments on multiple tracking benchmarks, including OTB2015, VOT2020, TrackingNet, GOT-10k, LaSOT, our proposed method named MGTT, the results could demonstrate that the proposed tracker achieves better performance than multiple state-of-the-art trackers.