Training Efficient Siamese Network with Probability Label for Object Tracking

Zanzhou Bai, Zihao Li, Xingqi Fang, Yueyang Gu, Jiyue Sun, Suwei Ma, Xiwen Zhang, Hang Xu, Yu Qiao

Abstract: Most object tracking algorithms based on siamese network are trained end-to-end with binary labels. However, the binary labels with only “1” and “0” states may not be a perfect way to represent the similarity between exemplar image and candidate in search image. In this paper, we analyze the weakness of training algorithm with binary labels and explore the scheme of utilizing probability labels to train a more efficient siamese network. The distribution of probability labels is yielded by an exemplary siamese network with state-of-art performance. The probability label indicates the probability of a candidate being the target, which provides more supervised information than binary label. With the probability labels, we attempt to design a more efficient siamese network. We conduct experiments on OTB-2013/2015 and VOT-2016/2017/2019, which demonstrate that training same siamese network with probability labels can achieve better performance than that with binary labels and other trackers.