Dual-Path Part-Level Method for Visible–Infrared Person Re-identification

Abstract

Visible–infrared cross-modality person re-identification is a realistic problem of person re-identification. Under poor illumination scenario, general methods of visible–visible person re-identification can not solve the problem well. If we directly compare the visible images of pedestrians captured under dark lighting with the visible images of pedestrians captured under normal light, this extreme color deviation will greatly reduce the recognition ability of the learned representations. In this paper, we propose a dual-path framework for visible–infrared cross-modality person re-identification based human part level features. Feature learning module contains modality-specific dual-path layers and modality-shared human part-level layers, which achieve discriminative global and local representations. In order to better optimize the proposed network, we design a global loss function and a local loss function for the global features and local features, respectively. The two loss functions are integrated together to train the network. We verify the effectiveness of our method on the challenging benchmarks: SYSU-MM01 and RegDB. Experimental results show that, compared with other cross-modality methods, our method has better effect in improving visible–infrared cross-modality person re-identification tasks.