Deep Multi-Labelling based Unsupervised Person Re-Identification

Abstract

Person re-identification is the problem of finding the given person in the nonoverlapping cameras at a given time or finding the person that has appeared in the same camera but at different time stamps. The re-identification problem plays vital role in the field of visual surveillance. In real world scenario, the problem is not as easy as it seems because of the various obstacles including but not limited to occlusion, lighting, viewpoint variation and low-quality data. Mostly techniques are presented in supervised manner, but they seem to perform well only in the case of abundant labelled data which does not make sense in real life scenario because it is hard to label huge amount of data which increases in every passage of time and it requires a great amount of not only resources but time as well. We present an unsupervised technique, that not only optimizes deep feature learning but also utilizes the significant information of auxiliary references other than visual feature similarity in large-scale re-Id.