I. Introduction
person re-identification (re-ID) aims to retrieve pedestrians with the same identity as the given probe image across non-overlapping camera views. To meet increasing demands in urban security, person re-ID is an essential intelligent video surveillance technique. Relying on video surveillance systems widely deployed in railway stations, subway stations, supermarkets, and other crowded public places, person re-ID technology is of significant importance in improving the efficiency of tracing criminal suspects, crime prevention, and missing population investigation. Traditional fully-supervised works achieve impressive performance, where all identity annotations are available during training [1], [2], [3], [4]. Such methods, however, heavily rely on manually-labelled person identity annotations, which are pretty costly to collect. Thus, it becomes formidable to apply these fully-supervised methods into large-scale practical scenarios, where many person images need to be annotated in a limited time. To address this issue, the unsupervised re-ID methods, without the need for identity annotations during training, have been studied by an increasing number of researchers [5], [6], [7], [8].