DiffusionTracker: Targets Denoising Based on Diffusion Model for Visual Tracking

The problem of background clutter (BC) is caused by distractors in the background that resemble the target’s appearance, thereby reducing the precision of visual trackers. We consider these similar distractors as noise and formulate a denoising task to solve the visual tracking problem. We propose a target denoising method based on a diffusion model for visual tracking, referred to as DiffusionTracker, which introduces the diffusion model to distinguish between targets and noise (distractors). Specifically, we introduce a reverse diffusion model to eliminate noisy distractors from the proposal candidates generated by the Siamese tracking backbone. To handle the difficulty that distractors do not strictly conform to a Gaussian distribution, we incorporate Spatial-Temporal Weighting (STW) to integrate spatial correlation and noise decay time information, mitigating the impact of noise distribution on denoising effectiveness. Experimental results demonstrate the effectiveness of the proposed method, with DiffusionTracker achieving a precision of 64.0% on BC sequences and a success rate of 63.8% on BC sequences from the LaSOT test datasets, representing improvements of 11.7% and 10.2% respectively over state-of-the-art trackers. Furthermore, our proposed method can be seamlessly integrated as a plug-and-play module with cutting-edge tracking algorithms, significantly improving the success rate for tracking task in background clutter scenarios.