Unsupervised person re-identification via camera-aware multi-level label refinement.

Abstract

Unsupervised person re-identification (re-ID) aims to match individuals across camera views without manual annotations, making it a challenging yet promising task. Although recent methods have made notable progress by leveraging pseudo-labels, two key challenges remain insufficiently addressed: (1) the inherent noise in pseudo-labels stemming from clustering, and (2) the limited discriminability of features resulting from camera variation. To address these issues, we propose a camera-aware multi-level label refinement (CMLR) framework, which jointly refines labels at both cluster and instance levels to facilitate more effective contrastive learning and enhance feature discrimination. At the cluster level, our dual-level intra-inter refinement (DIIR) module exploits intra- and inter-camera relationships to improve global and local pseudo-labels. At the instance level, the affinity-guided mutual refinement (AGMR) module computes affinity scores between samples based on selected informative nodes, adaptively pulling reliable positive pairs closer while pushing negative ones apart. By integrating camera-aware cues into multi-level refinement, CMLR enhances intra-class cohesion and inter-class separation, enabling more robust feature learning. Extensive experiments on Market-1501 and MSMT17 demonstrate the superiority of our CMLR compared to state-of-the-art unsupervised re-ID approaches.