Abstract How to effectively exploit spatio-temporal information is crucial for accurate target localisation in visual tracking. However, most existing trackers attempt to devise complex appearance models or template updating strategies, while neglecting the mining of the relationship between the target and the background. To address these issues, we propose a novel target-background interaction prompt framework for visual tracking, called TBPTrack. Firstly, we design a multi-scale target-background feature extraction module, which can fully excavate the features of target-background on more and applicable scales, and relax the influence caused by the appearance change of target. Secondly, a target-background prompt generator is presented, which utilises target-background tokens to generate explicit visual prompts that promote inferences in the current frame. The introduction of background tokens can not only enhance the representation of targets, but also weaken the interference of similar objects inter-class. Finally, a spatio-temporal scenario module is designed, utilising spatio-temporal tokens to propagate information between consecutive frames without the necessity of updating templates. This alleviates the challenge of determining the optimal moment for template updating. Extensive experimentation in seven benchmark tests demonstrates the effectiveness of the proposed method in comparison to existing state-of-the-art trackers.
Zhang et al. (Tue,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: