基于伪深度特征自蒸馏的域泛化行人重识别OA

Depth information provides a complementary modality in person re-identification(ReID),which effectively alleviates the over-reliance on texture features and enhances the model's generalization ability in cross-domain scenarios.Motivated by the observation that local similarity facilitates learning domain-invariant representations,we propose a depth-guided self-distillation for domain generalization framework.Specifically,we leverage Depth Anything to generate pseudo depth maps as cross-domain consistency self-su-pervised signals,and design a pseudo-supervised depth feature extraction mechanism with dual-dimensional attention to enable geometry-aware representation learning.Furthermore,a cross-domain depth similarity module and an edge similarity module are introduced to achieve geometry-guided cross-domain feature disentanglement.To enhance domain invariance,we construct a dynamic memory bank to store depth,edge,and local features,and adopt a dual-domain reciprocal self-attention mechanism to mine semantic cues that are or-thogonal to identity classification through contrastive learning.Ultimately,the proposed framework transforms geometric consistency constraints into implicit regularization for the classification task,thereby improving generalization while preserving discriminative power.Extensive experiments on benchmark cross-domain datasets,including Market1501,MSMT17,CUHK-SYSU,CUHK03-NP,and RandPerson,demonstrate the effectiveness and superiority of the proposed framework.