Abstract
Visible-infrared person re-identification (VI-ReID) remains a challenging task due to significant cross-modal discrepancies and poor image quality. While existing methods predominantly employ deep and complex neural networks to extract shared cross-modal features, these approaches inevitably discard critical primitive features during high-level feature abstraction. To address this limitation, we propose the High-order Interaction and Wavelet Convolution Network (HIW-Net) that systematically integrates primitive features at multiple feature interaction stages, thereby compensating for information loss in High-order representations. Furthermore, our framework uses wavelet convolution to mine more diverse features and solve the problem of insufficient feature extraction. We create the RegDB_shape datasets with the help of the Segment Anything Model(SAM) tool to supplement the training set. Extensive experiments on the SYSU-MM01 and RegDB datasets show the superiority of the proposed HIW-Net over several other state-of-the-art methods, proves the effectiveness of this method.