Survey on Unsupervised Techniques for Person Re-Identification

Abstract

Unsupervised person re-identification (ReID) might be difficult if lacking labeling information. The feature extraction scheme generally divides existing methods into handcrafted feature-based methods, unsupervised domain adaptation (UDA) based methods, and pseudo-labels estimation-based methods. Feature representations are extracted or learnt directly from unlabeled datasets to address the scalability issue by hand-crafted feature-based methods. The purpose of unsupervised domain adaptation is to relieve the domain bias as the learnt features are transferred to an unlabeled target from a labeled source. For pseudo-labels estimation-based methods, they take supervised pseudo-labels to learn feature representations and labels are estimated together for unlabeled datasets. In this paper, the state-of-the-art unsupervised techniques are reviewed to solve the task of person re-identification, a brief review of each method along with their evaluations on a set of widely used datasets in included. In addition, we give a detail comparison among these methods according to corresponding category.