SIST: Online Scale-Adaptive Object tracking with Stepwise Insight

Abstract

Correlation filter (CF) has achieved great success in the tracking community due to its high efficiency and effectiveness. However, in most existing CF trackers: (i) they mainly focus on object localization and cannot handle scale variations well; (ii) the learned CF is unavoidably deteriorated by the boundary effects, especially the background pixels inside the bounding box; (iii) there exists a localization gap between the original image and feature map. To handle these issues, we propose an online scale-adaptive correlation tracker by collaboratively learning a target-insight model (TIM) for object localization and a scale-insight model (SIM) for progressive refinement. In TIM, we introduce a target likelihood map to impose discriminative weights on the image, and then formulate a target-insight correlation filter for encoding the target as well as its surrounding context, which can handle boundary effect and distractors. Furthermore, we developed a SIM based on multiple instance learning and structured-labelled samples, which realizes the scale estimation and position refinement over time. Specifically, a hard negatives mining strategy is introduced to update the SIM rather than random sampling, which helps realize long-term tracking with re-capturing the missing target even tracking failure occurs. Extensive experiments on four benchmarks demonstrate that our method outperforms several state-of-the-art methods.