Contrastive-Siamese Collaborative Network for Visual Object Tracking
Ning Sun, Zhibin Zhang, Wanli Xue · 2022
The siamese network-based tracking framework aims to use the features extracted by the convolutional network to discriminate the similarity between the search image and the target template at the pixel level. This approach prevents the tracker from accurately tracking targets when dealing with tracking challenges such as analogue interference, fast movement, and occlusion. In this paper, we propose a novel tracking framework based on contrastive learning and the collaboration of siamese networks. In this framework, a kind of multi-task learning is formed by combining training with comparison learning task and tracking task. In the training process, the tracker can not only distinguish the target from the background but also learn the essential characteristics of the target without paying attention to the details of the pixel level through the contrastive learning method, to effectively distinguish the target from similar objects in complex scenes. Compared with other methods on three standard datasets OTBI00, VOT2015 and TColor128, our algorithm achieves better performance.