Higher-order moment alignment with uncertainty awareness for visible-infrared person re-identification

Visible-Infrared Person Re-Identification (VI-ReID) encounters substantial challenges from the inherent modality gap, with existing methods predominantly focusing on second-order statistics while overlooking higher-order distributional properties. This study proposes an Uncertainty-Aware Higher-Order Moment Alignment (UA-HOMA) framework to explicitly model and align third-order (skewness) and fourth-order (kurtosis) moments between cross-modal feature distributions. The framework employs kernelized moment matching in Reproducing Kernel Hilbert Space to avoid expensive tensor operations, while incorporating variational dropout-based uncertainty quantification to dynamically weight the alignment process. The module integrates with transformer-based architectures for end-to-end optimization. Extensive experiments on SYSU-MM01 and RegDB datasets demonstrate that UA-HOMA achieves 67.5±1.1% mAP on SYSU-MM01, outperforming state-of-the-art methods by 1.3%. The higher-order moment alignment contributes an additional 2.2% mAP improvement over second-order baselines, while uncertainty weighting provides 0.7% further enhancement under challenging conditions. By unifying higher-order statistics with uncertainty-aware learning, this work establishes an effective approach for cross-modality person re-identification, offering practical improvements for surveillance systems under heterogeneous imaging conditions.