Unsupervised Person Re-Identification via Multi-Label Classification

The challenge of unsupervised person re-identification (ReID) lies in learning discriminative features without true labels. Most of previous works predict single-class pseudo labels through clustering. To improve the quality of generated pseudo labels, this paper formulates unsupervised person ReID as a multi-label classification task to progressively seek true labels. Our method starts by assigning each person image with a single-class label, then evolves to multi-label classification by leveraging the updated ReID model for label prediction. We first investigate the effect of precision and recall rates of pseudo labels to the ReID accuracy. This study motivates the Clustering-guided Multi-class Label Prediction (CMLP), which adopts clustering and cycle consistency to ensure high recall rate and reasonably good precision rate in pseudo labels. To boost the unsupervised learning efficiency, we further propose the Memory-based Multi-label Classification Loss (MMCL). MMCL works with memory-based non-parametric classifier and integrates local loss and global loss to seek high optimization efficiency. CMLP and MMCL work iteratively and substantially boost the ReID performance. Experiments on several large-scale person ReID datasets demonstrate the superiority of our method in unsupervised person ReID. For instance, with fully unsupervised setting we achieve rank-1 accuracy of 90.1% on Market-1501 , already outperforming many transfer learning and supervised learning methods.