Deep Constraints Space of Medium Modality for RGB-Infrared Person Re-identification

Reducing the gap between modalities is key to RGB-Infrared cross-modality person re-identification. In this paper, we propose an architecture based on the Deep Constrains Space of Medium Modality (DCSMM) for RGB-Infrared person re-identification. Specifically, a Medium Modality Network (MMN) is proposed to extract fused features of RGB and grayscale images, and we combine the fused features with infrared features for constraint. In addition, we also propose a loss function termed Domain Alignment and ID Consistency Loss (DAIC), which constrains the differences between the medium modality and the infrared modality as well as within single-modality in terms of instance level. Finally, in the high-level semantic stage, we also propose a Spatial Barycenter Margin Loss (SBM) based on each identity barycenter to constrain the feature space with different identities. The proposed method is validated on two large-scale datasets SYSU-MM01 and RegDB for cross-modality person re-identification, the results show that it achieves superior performance compared with the state-of-the-art methods.