G<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.svg"><mml:msup><mml:mrow /><mml:mn>2</mml:mn></mml:msup></mml:math>DA: Geometry-guided dual-alignment learning for RGB-infrared person re-identification

RGB-Infrared (IR) person re-identification aims to retrieve person-of-interest from heterogeneous cameras, easily suffering large image modality discrepancy caused by different sensing wavelength ranges. Existing works usually minimize such aligning distribution of global features, while neglecting deep semantics and high-order structural relations within each class. This might render the misalignment between samples. In this paper, we propose Geometry-Guided Dual-Alignment (G2DA) learning, which yields better sample-level alignment for RGB-IR ReID solving a graph-enabled matching task that maximizes agreement multi-modality node representations considering edge topology. Specifically, covert RGB/IR images into semantic-aligned graphs, in whole-part features their similarities are represented nodes associated edges, respectively. To simultaneously implement node- edge-wise (Dual Alignment), introduce Optimal Transport (OT) as metric calculate cross-modality human body scores. By minimizing displacement cost across G2DA could learn not just modality-invariant but structurally consistent representations. Furthermore, advance Message Fusion Attention (MFA) mechanism adaptively smooth graph, effectively alleviating occlusions other individuals and/or objects. Extensive experiments on two standard benchmark datasets validate superiority G2DA, yielding competitive performance against previous state-of-the-arts.