Visible-infrared cross-modality person re-identification based on whole-individual training

摘要

Visible-infrared cross-modality person re-identification (VI-ReID) aims to search person images across cameras of different modalities, which can make up for the problem that ReID cannot be performed through visible images in a dark environment. The difficulty of VI-ReID task is the huge discrepancy between the visible modality and the infrared modality. In this paper, a novel whole-individual training (WIT) model is proposed for VI-ReID, which is based on the idea of pulling in the whole and distinguishing the individuals. Specifically, the model is divided into a whole part and an individual part. Two loss functions are developed in the whole part, namely center maximum mean discrepancy (CMMD) loss and intra-class heterogeneous center (ICHC) loss. Ignoring identity difference and treating each modality as a whole, the CMMD loss pulls in the centers of the two modalities. Ignoring modality difference and treating each identify as a whole, the ICHC loss pulls images with the same identity to its cross-modality center. In the individual part, a cross-modality triplet (CMT) loss is employed, which can distinguish the pedestrian images with different identities. The WIT model can help the network identify pedestrian images in an all-round way. Experiments show that the VI-ReID performance of the proposed method is better than existing technologies on two most popular benchmark datasets SYSU-MM01 and RegDB.(c) 2021 Elsevier B.V. All rights reserved.