Fully Unsupervised Person Re-Identification via Centroids and Neighborhoods Joint Learning

Abstract

This paper considers that the challenge of un-supervised person re-identification (re-ID) is generating high-quality pseudo labels. Recent label prediction methods can be mainly divided into Clustering-based Label Prediction (C-LP) and Similarity Measurements-based Label Prediction (SM-LP) methods. The existing researches only focus on improving the accuracy of one of the label generation method. In this letter, we first point out three complementarities between C- LP and SM-LP, including (1) interval of the pseudo label prediction (2) feature learning directions, and (3) inliers and outliers processing. Based on these three complementarities, we proposed a Joint Label Prediction (Joint-LP) method that can give full play to complementary advantages of C-LP and SM-LP. Moreover, we discover that standard Binary Cross Entropy (BCE) loss forces the unsupervised model to overfit the noisy labels, thereby leading the model training to fail. Therefore, we further proposed a Rectified Binary Cross Entropy (ReBCE) loss that is robust to label noise. The experimental results confirm the effectiveness of the proposed Joint-LP and ReBCE loss on two mainstream person re-ID datasets, Market-1501 and DukeMTMC-reID.