JROTM: Jointly Reinforced Object Tracking with Temporal Content Reference and Motion Guidance

Abstract Although the existing visual object tracking methods have made noticeable progress, many chokepoints remain unsettled. One of the most significant performance bottlenecks is the drastic change of the target, such as fast motion, scale variation, appearance changes, etc. Focusing on this major barrier, we propose our object tracking method, JROTM. It adopts the two-stage tracking-by-detection strategy, and it is reinforced by the temporal content reference (TCR) module and the motion guidance (MG) module. TCR precisely provides auxiliary frame-wise information of two consecutive frames of a video. With this information, the rapid change of the object’s appearance and position is retarded. MG module accurately offers more target position hints at the current frame. Our proposed JROTM is evaluated in multiple popular tracking benchmark, including OTB-100, TempleColor-128, and VOT-2016. It outperforms the state-of-the-art tracking methods by a non-neglectable margin.