cvpr18-Transferable Joint Attribute-Identity Deep Learning for Unsupervised Person Re-Identification

《cvpr18-Transferable Joint Attribute-Identity Deep Learning for Unsupervised Person Re-Identification》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、TransferableJointAttribute-IdentityDeepLearningforUnsupervisedPersonRe-IdentificationJingyaWang1XiatianZhu2ShaogangGong1WeiLi1QueenMaryUniversityofLondon1VisionSemanticsLtd.2{jingya.wang,s.gong,w.li}@qmul.ac.ukeddy@visionsemantics.comAbstractinthereal-worldasthereareaquadraticnumberofcamerap

2、airs,butalsoimplausibleinmanycases,e.g.theremaynotMostexistingpersonre-identification(re-id)methodsre-existsufficienttrainingpeoplereappearingineverypairofquiresupervisedmodellearningfromaseparatelargesetcameraviews.Thisscalabilitylimitationseverelyreducesofpairwiselabelledtrainingdataforever

3、ysinglecameratheusabilityofexistingsupervisedre-idmethods.pair.Thissignificantlylimitstheirscalabilityandusabil-Onegenericsolutiontolargescalere-idinreal-worldityinreal-worldlargescaledeploymentswiththeneedfordeploymentisdesigningunsupervisedmodels.Whileafewperformingre-idacrossmanycameravie

4、ws.Toaddressunsupervisedmethodshavebeendeveloped[13,9,21,20,thisscalabilityproblem,wedevelopanoveldeeplearn-34,51,61],theytypicallyofferweakerre-idperformancesingmethodfortransferringthelabelledinformationofanwhencomparedtothesupervisedcounterparts.Thismakesexistingdatasettoanewunseen(unlab

5、elled)targetdo-themlessusefulinpractice.Onemainreasonisthatwith-mainforpersonre-idwithoutanysupervisedlearninginoutlabelleddataacrossviews,unsupervisedmethodslackthetargetdomain.Specifically,weintroduceanTransfer-thenecessaryknowledgeonhowvisualappearanceofiden-ableJointAttribute-IdentityDee

6、pLearning(TJ-AIDL)forticalobjectschangescross-viewsduetodifferentviewan-simultaneouslylearninganattribute-semanticandidentity-gles,backgroundandillumination.Anothersolutionistodiscriminativefeaturerepresentationspacetransferrabletoexploitsimultaneously(1)unlabelleddatafromatargetdo-anynew(u

7、nseen)targetdomainforre-idtaskswithoutthemainand(2)existinglabelleddatasetsfromsometrainingneedforcollectingnewlabelledtrainingdatafromthetar-sourcedomains.Specifically,theideaistolearnafeaturegetdomain(i.e.unsupervisedlearninginthetargetdo-representation