基于交替中心聚类的无监督跨模态行人重识别OA

In addressing the challenge that existing unsupervised cross-modal person re-identification methods typically performed independent clustering for each modality,making it difficult to effectively establish cross-modal associations,an alternating center clustering strategy and optimization method was proposed.First,the data of different modalities were clustered independently to form initial clusters within each modality.Then,using the prior information of these initial clusters,an alternating approach was applied,where the centers of different modality clusters were selected as global initial cluster centers for secondary clustering,thereby constructing cross-modal cluster associations.Additionally,a hybrid contrastive learning framework was designed to further reduce the modality discrepancies by jointly learning modality-specific and modality-invariant features through the optimization of cross-modal feature representations with hybrid clustering centers.Experimental results show that compared to the baseline model,the proposed model achieves a performance improvement in mAP of 7.09％and 5.89％on the SYSU-MM01 dataset,and 18.52％and 18.30％on the RegDB dataset under two modes,respectively.