Dual-Path Model for Person Re-Identification Under Cloth Changing

Abstract

Most of the existing person re-identification (ReID) methods relies heavily on a person's clothes since clothing information is the clear and remarkable visual feature when the face of a person is unclear. However, in reality, people does not always wear the same cloth across camera views. Even worse, an adversary may change the clothes aiming to evade the identification. Some studies confirms that clothes changing downgrades the existing ReID methods significantly. The current ReID method considering clothes-changing does not fully utilize the person discriminant features, which may reduce its accuracy. This paper presents a dual-path model to learn the robust features under clothes changing and also the discriminant features for ReID from a RGB image and its contour sketch image respectively. The appearance and shape features of a person extracted by the two branches of our model are then combined to make a decision. The clothing information is eliminated from the appearance features by encouraging the similarity between the learned appearance and shape features. The experimental results on the PRCC dataset demonstrate that our model achieves higher performance under clothes changing compared to state-of-the-art ReID methods.