Attention fusion and target-uncertain detection for discriminative tracking

Abstract

Most existing prevalent trackers draw support from deep learning networks to acquire more effective target features, while we found that the dominant discriminative trackers employ a plain convolution block to turn backbone features into target classification features suitable for tracking. In this paper, we propose a lightweight feature separation and fusion module to obtain more effective and efficient semantic features in an end-to-end manner. Moreover, discriminative trackers entail collecting a certain number of new samples to online optimize the target classifier for retaining its discriminative ability, but it is not easy to pick out a more reliable sample for storage. Therefore, a feasible target-uncertain detection technique is devised to alleviate the tracking model corruption problem. In order to demonstrate the strong effectiveness and compatibility of our proposed approaches, we choose the excellent SuperDiMP and ToMP as baseline methods and conduct comprehensive experimental evaluations on seven public benchmarks. The results reveal that our methods perform superiorly against several state-of-the-art trackers on challenging benchmarks. The code and trained models will be available at https://github.com/hexdjx/VisTrack.