Updating Siamese trackers using peculiar mixup

Siamese network-based trackers are commonly used for tracking due to their balanced accuracy and speed. However, the fixed template from the initial frame is used to localize the target in the current frame, and never being updated in Siamese trackers. This loses the power of updating the appearance model online and easily induces tracking failures for the intrinsic properties of tracking processes such as constantly changing scenes and endless distractors. To address this problem, we propose a simple yet effective visual tracking framework for updating the appearance model in a novel formulation that introduces a peculiar mixup method both for training and inference phase of Siamese trackers (named PMUM-Siam). It consists of a template matching network and a mixup module. First, instead of center-cropping the inexact predictions for tracking, we use the template matching network which trained with predefined anchor boxes to learn to select the best candidate from similar distractors. Second, the mixup module is used to fuse and update the trade-off appearance model between the best candidate and the groundtruth. Our method greatly enhances the capability of target identification and target localization in Siamese trackers. To further demonstrate the generality of the proposed method, we integrate our PMUM-Siam into two representative Siamese trackers (SiamFC and SiamRPN+ +). Extensive experimental results and comparisons on five challenging object tracking benchmarks including OTB-2013, OTB-2015, OTB-50, VOT-2016, and VOT-2018 show that PMUM-Siam achieves leading performance with an average speed of 300 FPS.