Unsupervised Person Re-identification via Mining Label Homogeneity

Abstract

This paper studies the fully unsupervised person reidentification (Re-ID) problem which does not need any labeled data to train the re-ID network. To make fully unsupervised training possible, the model needs to generate pseudo labels. Unlike human-labeled annotation, the generated pseudo labels contain noisy labels. One of the best fully unsupervised methods till now formulated person Re-ID as a multi-label classification task. This approach only considered similarities between images and ignore relations among the generated labels. In this paper, we follow the multi-label classification training strategy but utilize the generated labels as auxiliary labels to further refine the noisy labels. The proposed approach seeks auxiliary labels by investigating two underlying homogeneity in generated labels, i.e. Symmetric Homogeneity (S) and Neighbor Homogeneity (N). The network is optimized under the joint supervision of the main pseudo label and two auxiliary labels in our proposed mutual label learning training strategy. The experimental results confirm the effectiveness of our discovered label homogeneities and the proposed mutual label learning on two mainstream datasets, Market-1501 and DukeMTMC-reID.