Person Re-Identification via Feature Representation Learning Based on Verification Sample Constrain

Ruifeng Zhao, Ajian Liu, Yanyan Liang, Haozhi Huang

Abstract: In person re-identification (ReID) task, the variance between the samples are quite large, and there is no standard sample as a comparison. The current common method is to implement it as a classification task, which can get better results than the classical methods in verification task such as contrastive loss. However, the identification loss used in classification task only seeks the boundary of the classification, the intra-class distance between samples is still large so that it is insufficient for ReID task. In this paper, we consider to overcome these difficulties by proposing a joint loss with an identification loss and constrains of modeling the Euclidean distances between samples and their corresponding eigen data which is constructed by Principal Component Analysis method (PCA), we call it EigenPerson. The entire loss is formed in a linear combination. This work is mainly motivated by the center loss for face recognition problems, of which regularizations are restricted by a simultaneously learned center. We substitute the center with EigenPerson which we constructed offline as an auxiliary training sample. The learned model with our proposed method is evaluated on the benchmark of Market1501 and CUHK03 and achieve comparable results to those methods proposed in the same period.