Yuan Chao Zhang Guiwei Ma Changxiao Zhang Tianyi Niu Guanglin

Abstract

Person re-identification (ReID) aims to extract accurate identityrepresentation features. However, during feature extraction, individual samplesare inevitably affected by noise (background, occlusions, and modellimitations). Considering that features from the same identity follow a normaldistribution around identity centers after training, we propose a Training-FreeFeature Centralization ReID framework (Pose2ID) by aggregating the sameidentity features to reduce individual noise and enhance the stability ofidentity representation, which preserves the feature's original distributionfor following strategies such as re-ranking. Specifically, to obtain samples ofthe same identity, we introduce two components: Identity-Guided PedestrianGeneration: by leveraging identity features to guide the generation process, weobtain high-quality images with diverse poses, ensuring identity consistencyeven in complex scenarios such as infrared, and occlusion. Neighbor FeatureCentralization: it explores each sample's potential positive samples from itsneighborhood. Experiments demonstrate that our generative model exhibits stronggeneralization capabilities and maintains high identity consistency. With theFeature Centralization framework, we achieve impressive performance even withan ImageNet pre-trained model without ReID training, reaching mAP/Rank-1 of52.81/78.92 on Market1501. Moreover, our method sets new state-of-the-artresults across standard, cross-modality, and occluded ReID tasks, showcasingstrong adaptability.