Robust Long-Term Tracking via Instance-Specific Proposals

Abstract

Correlation filter-based trackers are able to achieve long-term tracking when an additional detector is available. However, it is still challenging to achieve robust and accurate tracking due to several complicated situations, including occlusion and severe deformation. This is because a simple model is difficult to adapt to dramatic appearance changes of the object. Furthermore, redetection results are sometimes unreliable as the detector is trained on only a limited number of samples. In this paper, we propose a tracker named Complementary Learners with Instance-specific Proposals (CLIP). This tracker consists of three major components, including a translation filter, a scale filter, and an error correction (EC) module. The translation filter incorporates complementary features to handle severe target appearance variations. It is further combined with the scale filter to predict the state of the target. Finally, an adaptive updating mechanism is proposed to balance the stability and flexibility of the tracking model. Moreover, an instance-specific proposal generator is embedded into the EC module to recover the lost target from tracking failures. The experimental results on OTB2015, VOT2016, Temple-Color 128, and UAV20L demonstrate that the proposed CLIP tracker achieves comparable performance with the state-of-the-art trackers on several scenarios, including occlusion, deformation, and out-of-plane rotation. Moreover, the CLIP tracker is able to run at a speed of 35 frames/s on OTB2015, which makes it highly suitable for many real-time applications.