Deep Transfer Learning for Person Re-identification英文文献资料

《Deep Transfer Learning for Person Re-identification英文文献资料》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、DeepTransferLearningforPersonRe-identificationMengyueGengYaoweiWangTaoXiangYonghongTianAbstractinCUHK03[25]and1,501inMarket1501[66].Incon-trast,thewidelyusedLFWdataset[17]forfaceverificationPersonre-identification(Re-ID)posesauniquechal-has5,749identities–facesofcelebritiesaremucheasierlengeto

2、deeplearning:howtolearnadeepmodelwithmil-tocollectandlabelthanpassers-bycapturedbyasurveil-lionsofparametersonasmalltrainingsetoffewornola-lancecameranetwork.Importantly,onecouldeasilycol-bels.Inthispaper,anumberofdeeptransferlearningmod-lectamuchlargerauxiliarydatasetoffacestoassistintheel

3、sareproposedtoaddressthedatasparsityproblem.First,modellearning:oneofthestate-of-the-artresultsonLFWadeepnetworkarchitectureisdesignedwhichdiffersfromwasobtainedbypretrainingthedeepmodelonanauxiliaryexistingdeepRe-IDmodelsinthat(a)itismoresuitablefacedatasetof200Mimagesof8Midentities[41].fo

4、rtransferringrepresentationslearnedfromlargeimageGiveninsufficienttrainingsamples,transferringfeatureclassificationdatasets,and(b)classificationlossandveri-representationslearnedfromalargerauxiliarydatasetbe-ficationlossarecombined,eachofwhichadoptsadifferentcomesnecessary.Indeed,transferlearni

5、nghasbeencon-dropoutstrategy.Second,atwo-steppedfine-tuningstrategysideredinmostexistingdeepRe-IDworks.Inparticular,isdevelopedtotransferknowledgefromauxiliarydatasets.givenasmallRe-IDdatasetwithonlyafewhundredsofThird,givenanunlabelledRe-IDdataset,anovelunsuper-labelledidentities,existingmo

6、delstypicallypretrainwithviseddeeptransferlearningmodelisdevelopedbasedonlargerRe-IDdatasetsfollowedbyfine-tuningonthetargetco-training.Theproposedmodelsoutperformthestate-of-set,withanotableexceptionof[55]whichlearnsasin-the-artdeepRe-IDmodelsbylargemargins:weachieveglemodeljointlyacrossmul

7、tipleRe-IDdatasetsbeforetheRank-1accuracyof85.4%,83.7%and56.3%onCUHK03,fine-tuningineach.Inotherwords,onlyRe-IDdatasetsareMarket1501,andVIPeRrespectively,whilstonVIPeR,ourconsideredasauxiliarydatasets–hardlyidealbecauseallunsupervisedmodel(45.1%)beatsmost