The online scene-adaptive tracker based on self-supervised learning

In recent years, with the advancement of deep learning technology, visual object tracking has developed rapidly. However, the real world is complex and changeable, and the CNN-based trackers lack an effective updating mechanism for unseen targets and scene adaptation. Aiming at the two key issues that existed in current trackers algorithms: (1) the poor generalization with the limited offline training data; (2) lacking an online model updating mechanism for scene adaptation, this paper designs a simple and effective self-supervised framework for online model adaptation and propose OSATracker, which construct a cycle self-supervision based on unlabeled online data. OSATracker utilizes decision-level information to update the model rather than template by self-supervised optimization in the online tracking process, which effectively improves the generalization ability of the model in the tracking process. Extensive experiments on visual tracking benchmarks including VOT2018, VOT2016, OTB2015, and OTB2013 demonstrate the effectiveness of OSATracker with outstanding performance.