A Discriminatively Learned CNN Embedding for Person Re-identification

《A Discriminatively Learned CNN Embedding for 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、ADiscriminativelyLearnedCNNEmbeddingforPersonRe-identificationZhedongZheng,LiangZheng,YiYangUniversityofTechnologySydney,Australia{zdzheng12,liangzheng06,yee.i.yang}@gmail.comAbstractWerevisittwopopularconvolutionalneuralnetworks(CNN)inpersonre-identification(re-ID),i.e.,verifi

2、cationandclassificationmodels.Thetwomodelshavetheirre-spectiveadvantagesandlimitationsduetodifferentlossfunctions.Inthispaper,weshedlightonhowtocombinethetwomodelstolearnmorediscriminativepedestriande-scriptors.Specifically,weproposeanewsiamesenetworkthatsimultaneouslycomputes

3、identificationlossandverifi-cationloss.Givenapairoftrainingimages,thenetworkpredictstheidentitiesofthetwoimagesandwhethertheybelongtothesameidentity.Ournetworklearnsadiscrimi-nativeembeddingandasimilaritymeasurementatthesametime,thusmakingfullusageoftheannotations1.Albeitsimpl

4、e,thelearnedembeddingimprovesthestate-of-the-artFigure1.Thedifferencesbetweenverificationandclassificationperformanceontwopublicpersonre-IDbenchmarks.Fur-models.GreenblocksrepresentnonlinearfunctionsbyCNN.a)ther,weshowourarchitecturecanalsobeappliedinimageClassificationmodelstr

5、eatpersonre-IDasamulti-classrecogni-retrieval.tiontask,whichtakeoneimageasinputandpredictitsidentity.b)Verificationmodelstreatpersonre-IDasabinary-classrecogni-tiontaskorsimilarityregressiontask,whichtakeapairofimages1.Introductionasinputanddeterminewhethertheybelongtothesame

6、personornot.Herewejustshowabinary-classrecognitioncase.Personre-identification(re-ID)isusuallyviewedasanimageretrievalproblem,whichmatchespedestriansfromdifferentcameras[33].Givenaperson-of-interest(query),abinary-classclassificationtaskorasimilarityregressionpersonre-IDdeterm

7、ineswhetherthepersonhasbeenob-arXiv:1611.05666v1[cs.CV]17Nov2016task[31,12,28,23].Givenalabels∈{0,1},theverifi-servedbyanothercamera.cationnetworkforcestwoimagesofthesamepersontobeRecently,convolutionalneuralnetworks(CNN)havemappedtonearbypointsinthefeaturespace.Iftheimagessh

8、ownpotentialforlearningstate-of-the-artfeatureembed-areofdifferentpeoplethe