Unsupervised Person Re-Identification based on Feature Learning of Deep Convolutional Generative Adversarial Network

Abstract

Most of the researchers are targeting the problem of person re-identification with supervised learning. Person re-identification problem requires a lot of pairwise manually annotated training data for each camera. This approach becomes impractical for large scale deployment because of the dependency on pairwise manually annotated data. In this paper, we propose an unsupervised approach for the person re-identification problem based on utilization of Generative Adversarial Network. We use Deep Convolutional Generative Adversarial Network for feature learning from image sequences of datasets as we prove that GANs learn the features effectively. In the feature learning process, we do not use any labels so this is a completely unsupervised process. Several similarity measures and a Cumulative Match Characteristic (CMC) Curve are used for the evaluation of our results. The experimental results show that our proposed approach gives comparable results to the existing state of the art approaches.