Learning Self-Corrective Network For Long-Term Object Tracking

Existing long-term trackers often online update their models to adapt target appearance variations,but the inaccurate tracking results introduce cumulative error into the updated model that may cause severe drift issue.To this end,a robust long-term tracker should have the self-corrective capability that can judge whether the tracking result is reliable or not,and then it is able to recapture the target when severe drift happens.To address how to judge the current tracking reliability,we design a Selfmodulated Tracking Reliability Evaluator(STRE)that is an effective tracking reliability classifier built on a modulation network.The classifier is trained through an end-to-end offline manner whose training samples are annotated from the unlabeled video with pseudo labels generated by an adaptive self-labeling strategy.The adaptive self-labeling can automatically label hard samples that are often neglected in existing trackers according to the statistical characteristics of target state,and the network modulation mechanism can guide the backbone network to learn more discriminative features without extra training data.To address recapture the target after severe drift,this paper further designs a Selfadjusting Proposal Post-Processor(SPPP),and uses STRE as a trigger,used to trigger SPPP.SPPP is composed of memory storage module and dynamic NMS strategy.The memory storage model records the status information of the target during stable tracking based on the evaluation results of STRE,and provides reliable judgment basis for the dynamic NMS,so that when the tracker drifts,the SPPP module can recapture the target in a timely and accurate manner.Finally,both the STRE and SPPP construct our Self-Corrective Long-term Tracker(SCLT).Extensive evaluations on four large-scale and challenging benchmark datasets including VOT2021 LT,Ox Uv ALT,TLP,and La SOT with more than 2,183,000 frames demonstrate superiority of the proposed SCLT to a variety of state-of-the-art long-term trackers in terms of all metrics.