<p>Recent advancements in deep learning have significantly advanced visible-infrared person re-identification (VI-ReID). However, most existing methods prioritize either intrinsic sample features or cross-modal alignment, often neglecting the structural relationships between samples during feature learning. This limitation hinders the effective exploitation of relation-preserving features, thereby overlooking critical identity-relevant information embedded in modality-specific characteristics. To address this gap, we propose a novel cross-modal Relation-preserving Feature Embedding (RFE) network, which improves the performance of VI-ReID models by leveraging structural correlations between samples across modalities, independent of external annotations. RFE integrates both the content of individual samples and their neighborhood relationships into the feature embedding process, effectively combining the advantages of relational and supervised contrastive learning. A key innovation is the relation-aware feature fusion module, which dynamically combines information from image representations and their component-level relationships to enhance relation-preserving feature embedding. This fusion process refines the learned features into shared cross-domain representations, improving their discriminative power. Experimental results on three VI-ReID benchmark datasets demonstrate the superiority of our method.</p>