Exploring Latent Information for Unsupervised Person Re-Identification by Discriminative Learning Networks

摘要

For unsupervised domain adaption in person re-identification (Re-ID) tasks, the generally used label estimation approaches simply use the global features or the uniform part features. They often neglect the variations of samples having the same identity caused by occlusion, misalignment and uncontrollable camera settings. In this paper, we propose a discriminative learning network with target domain latent information (LatentDLN) to enhance the generalization ability of the Re-ID model. Specifically, to generate a discriminative and robust representation, two types of latent information in the samples from the target domain are explored by the multi-branch deep structure. First, the key points based valid region information is used to leverage the local and global cues in human body, and then a heuristic distance metric learning method based on the global features and the local features is proposed to effectively evaluate the similarity between different images. Second, the camera style transferred images are used as augmentation data to bridge the gap between different cameras in target domains. Moreover, the re-rank mechanism based on reciprocal neighbors is designed to improve the quality of the label estimation. Experimental results on Market-1501, DukeMTMC-ReID and MSMT17 datasets validate the significant effectiveness of the proposed LatentDLN for unsupervised Re-ID.