Contrastive Learning with Information Compensation for Visible-Infrared Person Re-Identification

In the task of visible-infrared person re-identification (Re-ID), it is challenging to learn modality-invariant features, due to the significant modality gap and the absence of cross-modality matching pairs with identical content but different modalities. To address this issue, this paper proposes an information compensation based contrastive learning method for visible-infrared person Re-ID. Firstly, an information compensation mechanism based on an interme-diate modality is designed to achieve feature-level discrim-inability compensation and image-level cross-modality content-invariant compensation. The former leverages the sufficient samples and modality consistency in the intermediate modality, enabling its feature learning to focus more on effective identity recognition information. By integrating these features, it enhances the discriminability of the original visible and infrared modality features. The latter utilizes the formation of cross-modality matching pairs between visible-intermediate and infrared-intermediate modalities to enrich the diversity of samples for contrastive learning. Additionally, a multimodality contrastive learning loss is proposed to further improve the feature discriminability across different identities and enchance the robustness across modalities. Experimental results demonstrate the effectiveness of the proposed method.