Object Tracking in Videos: A Real-time Siamese Network with Online Update Module

One dominant tracking framework is the Siamese network, which uses the object from the first frame as template and find the best matched patch in every subsequent frame. Although showing good speed and accuracy trade-off, it lacks the capacity of adapting to rapid changing appearance. Even though some efforts have been made towards online update of template, they still suffer from resorting to less powerful classifiers or adding too heavy computation and storage burden. To tackle these problems, we propose a light-weight template updating module and embed it into the Siamese network for real-time visual tracking. In particular, this module continuously takes in the most recent tracked target together with the initial template to encode their correlation, which is then refined by message passing for effective online template adaption. On one hand, it dynamically captures the appearance change of the tracked target. On the other hand, it can selectively highlight the most discriminative representations in the template to preserve robustness. Then the updated template and the current candidate perform cross-correlation in a standard Siamese network for fast and accurate inference. We conduct extensive experiments to demonstrate the effectiveness of the proposed tracker, which surpasses our strong baseline and obtains competitive results on three large tracking benchmarks, VOT2018, LaSOT and VOT2016. Our tracking system runs at a speed of around 110 fps on one GTX Titan X GPU