Effective long-term tracking with contrast optimizer

The main challenge of long-term tracking includes data uncertainty in long-term observations. Previous methods tackle the long-term tracking task by online update-based trackers. However, sophisticated online update strategies of these trackers are usually with a considerable computational burden. In this work, a contrastive learning-based online optimizer-assisted long-term tracking framework (named LTCO) is proposed to guide the online tracker to make more accurate update decisions while reducing the impact of online updates on tracking speed. Specifically, the optimizer first perceives the similarity between distractors and positive samples through metric learning. Next, the contrastive learning between target anchors and hard negative samples forces the optimizer to notice the difference between targets and distractors. Finally, the optimizer will learn a binary output to assist the tracker updating. The proposed optimizer can be easily integrated into other online trackers with little impact on their running speed. Extensive experimental results show that the method achieves state-of-the-art performance on the VOT2018LT, VOT2019LT, OxUvA, and LaSOT benchmarks while running at real-time speed on GPU.