知识引导和细粒度信息增强的无监督域自适应行人再识别OA

Unsupervised domain adaptation(UDA)aims to transfer knowledge learned from a labeled source domain to an unlabeled target domain,playing a very important role in person re-identification.In real-world applications,video-based pedestrian data are often available,making it feasible to obtain single-camera-view labels in the target domain.However,existing UDA methods typically ignore this readily accessible information,thereby limiting performance improvements.To address this issue,we propose a knowledge-guided and fine-grained information enhancement framework for UDA person re-identification.A novel paradigm is introudced that leverages single-view labeled pedestrian samples in the target domain to fully exploit intra-domain information.Meanwhile,source-domain knowledge is used as guidance to assist the model to extract more discriminative target-domain pedestrian representations,effectively mitigating domain shift compared with conventional knowledge-transfer strategies.Furthermore,local pedestrian cues are integrated into global features to strengthen fine-grained feature expression.Experiments conducted on two publicly datasets fully demonstrate the effectiveness and superiority of the proposed method.