CMIA-Net: early cross-modal interaction for visible-infrared person re-identification

Abstract

Visible-infrared person re-identification (VI-ReID) remains challenging due to severe spectral discrepancy and local cross-modal misalignment between visible and infrared images. Most existing methods alleviate this discrepancy through middle- or late-stage feature alignment, but modality-specific shallow features may have already accumulated spectral bias and local correspondence errors before shared representations are formed. In this paper, we argue that VI-ReID should be treated as an early cross-modal correspondence learning problem rather than only a late embedding alignment problem. To this end, we propose CMIA-Net, a framework that establishes bidirectional visible-infrared interaction at shallow backbone stages. Its core module, cross-modal interaction attention (CMIA), enables visible and infrared feature maps to exchange complementary local information before deep semantic aggregation, thereby reducing progressive stream divergence. To stabilize early interaction, we further introduce spectral-invariant augmentation (MC-Aug) to suppress over-reliance on visible-spectrum cues and a bi-directional hetero-center learning (BHCT) loss to improve class-center-level cross-modal compactness and inter-class separability. Experiments on RegDB and SYSU-MM01 show that CMIA-Net achieves strong performance on RegDB and the SYSU-MM01 indoor-search protocol, while remaining competitive under the more challenging SYSU-MM01 all-search setting. Specifically, CMIA-Net obtains 94.09% Rank-1 accuracy and 90.74% mAP in the RegDB visible-to-thermal setting, and 66.67% Rank-1 accuracy and 58.12% mAP in the SYSU-MM01 all-search setting.