Part-guide teaching: part-guide pseudo-label self refinement for unsupervised person re-identification

Abstract

Unsupervised person re-identification (re-ID) aims to retrieve the same person across different camera-views for a given query without any labels. Recently, state-of-the-art methods that leverage pseudo-labels through clustering achieved competitive results. Despite improvements from the above methods, unreliable clusters and noisy pseudo-labels have remained problems. In this work, we propose Part-Guide Teaching (PGT) framework that effectively utilizes part-local features to address these problems. The proposed framework performs robust clustering based on part-wise relationships and refines hard pseudo-labels by guidance from part-local feature classifiers. Our refinement process works online in a self-teaching manner, which does not require additional teacher networks that need a high computational cost in the training stage. Experiment on Market-1501 and DukeMTMC-reID demonstrates the superiority of the proposed methods.