基于局部特征匹配和伪标签细化的纯无监督行人重识别

打开文本图片集

关键词：行人重识别；无监督；伪标签细化；局部特征匹配；神经网络；消融实验；相关性评分中图分类号：TN919.8-34；TP391 文献标识码：A 文章编号：1004-373X（2026）06-0174-10

Purely unsupervised person ReID based on local feature matching andpseudo-labelrefinement

LIU Guoquan12， CHEN Shangliang'，QIN Chenxu1， ZHOU Shumin1， ZHOU Huanyin1，WANG Xiaogang² （1.Schoolof MechanicalandElectronic Enginering，East China UniversityofTechnology，Nanchang 33013，China; 2.Artificial IntelligenceKeyLaboratory of SichuanProvince，Yibin644oO2，China）

Abstract：Inalusiontothe problemofsignificantnoise inpseudo-labels generatedbyclustering inusupervised personreidentification（ReID），apurelyunsupervisedmethodbasedonlocalfeaturematchingandpseudo-labelrefinementisproposed. This methoddoes notrelyonanysource domaininformation，butonlyconsiders thecorelation betweensamples attheimage levelandasignsrobustpseudo-labelsfortraining.Alocalfeaturematchingmoduleisdesignedtoalignandranklocalfeatures of samples，soastorepresentthecorelationbetweenglobalfeaturesandlocalfeaturesofsamplesreasoably.Then，acelation scoringmoduleisused toscoretherationalityofthegeneratedpseudo-labelsbyconsideringthecorrelation betweenglobal featuresandlocalfeaturescomprehensively.Onthisbasis，apseudo-labelrefinementmoduleisintroducedtorefinethepseudo labelsofglobalfeaturesandlocal featuresbasedonthescoresofsamples.Therefined pseudo-labelsareusedtotrainthenetworkandcontinuouslyupdatethepseudo-labels.Theexperimental verificationofthemethodisconductedonthepublicpersonReIDdatasetsMarket-1501，DukeMTMC-ReIDand MSMT17.TheresultsshowthatthemAPof thismethodcanreach 81.9% ， 71.1% and 31.6% on the Market-1501，DukeMTMC-ReID and MSMT17 datasets，respectively，demonstrating betr performance.

Keywords：personre-identification;unsupervised;pseudo-labelrefinement;localfeature matching;neuralnetwork;ablation experiment; correlation score

0 引言

视觉技术，旨在跨摄像头或跨场景条件下，对不同视角下的目标行人进行识别。（剩余21512字）