Enhancing Unsupervised Multi-Source Domain Adaptation for Person Re-Identification via Mixture of Experts and Graph-Based Relation

Person re-identification (re-ID) aims to match pedestrian images across disjoint camera views. Existing multi-source unsupervised domain adaptation (UDA) re-ID methods still face two critical issues: they fail to effectively balance domain-invariant feature learning and domain-specific style preservation and cannot adequately model the implicit correlations among diverse source domains, resulting in limited cross-domain generalization performance. To address these challenges, this paper proposes a novel multi-source UDA re-ID framework equipped with a Mixture of Experts feature extraction (MEFE) network and a Graph-Based Relation (GBR) module. Specifically, the MEFE network integrates mixed Instance and Batch Normalization (MIBN) to extract robust domain-invariant features, while the embedded domain-specific style information (DSI) module compensates for lost domain-specific style details at the feature level. Furthermore, the cascaded Graph Attention and Graph Convolution Networks (GATs/GCNs) in the GBR module adaptively explore implicit feature correlations and achieve effective multi-source feature fusion. Center maximum mean discrepancy loss is adopted to further reduce cross-domain distribution discrepancies. Extensive experiments on large-scale datasets demonstrate that the proposed method achieves state-of-the-art performance and substantially outperforms mainstream UDA re-ID approaches.