Multiple camera styles learning for unsupervised person re-identification

Abstract

Most clustering-based unsupervised person re-identification methods usually first assign pseudo-labels to pedestrian images using a clustering algorithm and then train a network using the labeled pedestrian images in each iteration. However, clustering algorithms may produce some clusters that only contain pedestrian images from fewer cameras, thus resulting in these clusters having fewer camera styles. Training a network using the pedestrian images from these clusters may degrade the capability of the network to retrieve pedestrians across cameras. Moreover, unsupervised learning methods usually utilize only contrastive loss that is based on self-supervised learning. However, lack of metric loss with mandatory guidance from labels may lead the network to fail to learn sufficiently the discriminative information of pedestrians. To address these issues, we propose a multiple camera styles learning (MCSL) method for unsupervised person re-identification. Specifically, we design a novel cluster filtration module (CFM) to ensure the clustering algorithm to keep the focus on pedestrian identities rather than camera styles by filtering out the clusters with fewer camera styles. In addition, we also design a multi-granularity feature metric (MGFM) loss to introduce metric learning with labels into unsupervised learning. The effectiveness of the proposed MCSL is demonstrated by the experimental results (Rank-1 accuracy) on Market1501 (93.7%), DukeMTMC(86.0%) and MSMT17 (71.12%), significantly reducing the gap between unsupervised and supervised performance on person re-identification.