SiamDMU: Siamese Dual Mask Update Network for Visual Object Tracking

The Siamese trackers have demonstrated powerful performance in object tracking, which convert tracking into template matching. However, the trackers may fail to track the target whose appearance changes drastically. The intuitive solution of the problem is to accurately update the template in time. After analyzing the previous works, we find that a good template update mechanism is desired to generate a comprehensive and robust representation, to grasp and exploit the target appearance change trend, and to be easy to train. Therefore, we propose a novel tracker named Siamese Dual Mask Update (SiamDMU), which utilizes long-term motion and semantic information to generate the enhanced tracking results for template update. SiamDMU exploits the long-term motion information of target at certain time intervals to capture the trend of target appearance change. The core structure of template update in SiamDMU consists of two convolution layers and be easy to train with 19 videos. Experiments on VOT2016, VOT2018, OTB100, UAV123 and LaSOT, show that SiamDMU has competitive performance with state-of-the-art trackers with real-time speed.