Pseudo Labels Refinement with Stable Cluster Reconstruction for Unsupervised Re-identification

Most existing unsupervised re-identification uses a clustering-based approach to generate pseudo-labels as supervised signals, allowing deep neural networks to learn discriminative representations without annotations. However, drawbacks in clustering algorithms and the absence of discriminatory ability early in training limit better performance seriously. A severe problem arises from path dependency, wherein noisy samples rarely have a chance to escape from their assigned clusters during iterative training. To tackle this challenge, we propose a novel label refinement strategy based on the stable cluster reconstruction. Our approach contains two modules, the stable cluster reconstruction (SCR) module and the similarity recalculate (SR) module. It reconstructs more stable clusters and re-evaluates the relationship between samples and clearer cluster representatives, providing complementary information for pseudo labels at the instance level. Our proposed approach effectively improves unsupervised reID performance, achieving state-of-the-art performance on four benchmark datasets. Specifically, our method achieves 46.0% and 39.1% mAP on the challenging dataset VeRi776 and MSMT17.