Continuous epoch distance integration for unsupervised person re-identification

Abstract

Unsupervised person re-identification aims to learn discriminative feature representations for person retrieval from unlabeled datasets. Clustering-based methods achieve state-of-the-art performance in this research direction. However, the noisy pseudo-labels generated during the clustering process and the interference caused by background noise limit the performance of the model. To this end, this paper proposes continuous epoch distance integration for unsupervised person re-identification, which performs clustering by integrating the distance matrix between two consecutive epochs to generate reliable pseudo-labels. In addition, the attention module of spatial structure and channel dimension is also proposed, so that the model pays more attention to the person itself and eliminates the interference of the background. The effectiveness of this method is verified on Market-1501, DukeMTMC-ReID, and MSMT17. The experimental results show that this method is superior to the current mainstream unsupervised person re-identification method.