ADTrack: Advanced Mamba fusion and distillation for single object tracking

Single Object Tracking is a critical task in computer vision, yet existing deep learning methods struggle with three major limitations: insufficient transfer of rich spatio-temporal relationships, the lack of parallel acceleration in State Space Model trackers like Mamba, and the inherent trade-off between speed and accuracy constrained by feature abstraction. To address these challenges, we propose ADTrack, a novel high-performance tracking framework. Specifically, ADTrack introduces an implicit state branch to facilitate the propagation of richer spatio-temporal information, significantly boosting tracking precision. Furthermore, we design the Mamba2-Equivalent Feature Fusion Neck, which leverages dual equivalence to achieve strong parallel acceleration capabilities for the SSM in the tracking pipeline. Finally, a novel self-distillation strategy is employed to enhance the information abstraction and extraction capability of the model. Extensive experiments on multiple challenging datasets, including TrackingNet, LaSOT, and GOT-10k, demonstrate that ADTrack achieves State-of-the-Art performance, especially when deployed at smaller model scales. Our framework effectively advances the performance boundary for contemporary visual tracking. A demo video is available at https://www.scidb.cn/anonymous/eWlNbmV1 .