Adaptive sparse attention-based compact transformer for object tracking

The Transformer-based Siamese networks have excelled in the field of object tracking. Nevertheless, a notable limitation persists in their reliance on ResNet as backbone, which lacks the capacity to effectively capture global information and exhibits constraints in feature representation. Furthermore, these trackers struggle to effectively attend to target-relevant information within the search region using multi-head self-attention (MSA). Additionally, they are prone to robustness challenges during online tracking and tend to exhibit significant model complexity. To address these limitations, We propose a novel tracker named ASACTT, which includes a backbone network, feature fusion network and prediction head. First, we improve the Swin-Transformer-Tiny to enhance its global information extraction capabilities. Second, we propose an adaptive sparse attention (ASA) to focus on target-specific details within the search region. Third, we leverage position encoding and historical candidate data to develop a dynamic template updater (DTU), which ensures the preservation of the initial frame's integrity while gracefully adapting to variations in the target's appearance. Finally, we optimize the network model to maintain accuracy while minimizing complexity. To verify the effectiveness of our proposed tracker, ASACTT, experiments on five benchmark datasets demonstrated that the proposed tracker was highly comparable to other state-of-the-art methods. Notably, in the GOT-10K 1 evaluation, our tracker achieved an outstanding success score of 75.3% at 36 FPS, significantly surpassing other trackers with comparable model parameters.