Occlusion Reconstruction for Person Re-identification

Abstract

Person re-identification is very important for monitoring and tracking crowd movement to provide security in public places. Although this domain has gotten enough attention in the last few years, with the introduction of deep learning to develop an automated re-identification model. There are several challenges in the task, one of them is performing re-identification in the presence of an unconstrained situation (i.e., occlusion) that has not been addressed effectively. In this work, we focus on re-identification from RGB frames captured by multiple cameras. Occlusion reconstruction in these captured non-sequential frames has been done by employing the generalization capability of generative adversarial networks (GANs). Our algorithm is based on a multi-model approach to jointly perform occlusion reconstruction along with re-identification. The performance of the proposed approach is evaluated on three publicly available datasets, and the results obtained are quite satisfactory and superior to all the existing approaches used in the study.