Unsupervised Person Re-identification with Multi-Level Feature Contrastive Learning

Abstract

Unsupervised person re-identification works mainly rely on feature representation learning. In recent years, many methods have used pseudo-labels generated from clustering and applied contrast learning techniques to train models. However, the existing methods only consider intra-class relationships in the mini-batch data, ignoring the relationships between classes, and can not effectively increase the distance between different pedestrian images. In addition, the use of a single feature vector does not fully utilize the information of all instance samples, and cannot represent the class features well. To address this issue, we propose an unsupervised multi-level feature contrastive learning framework for person re-identification. Specifically, the framework combines the contrastive learning loss of the farthest sample from the intra-class relationship, the farthest sample from the inter-class relationship and the cluster centroid contrastive learning loss for the unsupervised person re-identification model, fully mining more discriminative features related to pedestrian identity, and more effectively increasing the separability between classes and the similarity within the class. At the same time, combined with the generated data image as a enhancement dataset to join the network training, increase the diversity of samples, and effectively improve the robustness of the model. The effectiveness of our model is proved on the widely used re-identification dataset.