A Simple Similarity-Ranking-Based Pseudo-label Redistribution Method for Unsupervised Person Re-Identification

This is a pseudo-label refinement method for unsupervised Person ReID. The method select cluster sample that is most similar in a cluster to represent the cluster, and reassign pseudo-labels for those not most familiar with their corresponding cluster sample. Comprehensive experiments demonstrate that our method not only improves the external validity metric scores of the pseudo-labels, effectively narrowing the gap between pseudo-labels and the true distribution, and minimizing the accumulation of noisy label errors, but also significantly improves the performance of IICS on three public datasets – Market-1501, DukeMTMC-ReID, and MSMT17.  Our baseline method is IICS, and it may also improve other baseline such as PPLR. Future works will focus on the adaptive to all other baselines.