[CVPR 2013] Unsupervised Salience Learning for Person Re-identification

《[CVPR 2013] Unsupervised Salience Learning 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、UnsupervisedSalienceLearningforPersonRe-identificationRuiZhaoWanliOuyangXiaogangWangDepartmentofElectronicEngineering,TheChineseUniversityofHongKong{rzhao,wlouyang,xgwang}@ee.cuhk.edu.hkAbstractHumaneyescanrecognizepersonidentitiesbasedonsomesmallsalientregions

2、.However,suchvaluablesalientinformationisoftenhiddenwhencomputingsimilaritiesofimageswithexistingapproaches.Moreover,manyexist-ingapproacheslearndiscriminativefeaturesandhandledrasticviewpointchangeinasupervisedwayandrequirelabelingnewtrainingdataforadifferent

3、pairofcameraviews.Inthispaper,weproposeanovelperspectiveforper-sonre-identificationbasedonunsupervisedsaliencelearn-ing.Distinctivefeaturesareextractedwithoutrequiringidentitylabelsinthetrainingprocedure.First,weapplyadjacencyconstrainedpatchmatchingtobuilddens

4、ecor-respondencebetweenimagepairs,whichshowseffective-nessinhandlingmisalignmentcausedbylargeviewpointandposevariations.Second,welearnhumansalienceinanunsupervisedmanner.Toimprovetheperformanceofpersonre-identification,humansalienceisincorporatedin(a1)(a2)(a3)(

5、(a4)(a5)b5)(b4)(b3)((b2)b1)patchmatchingtofindreliableanddiscriminativematchedpatches.TheeffectivenessofourapproachisvalidatedonFigure1.ExamplesofhumanimagematchingandsaliencethewidelyusedVIPeRdatasetandETHZdataset.maps.Imagesontheleftoftheverticaldashedblackli

6、nearefromcameraviewAandthoseontherightarefromcameraviewB.Upperpartofthefigureshowsanexampleofmatchingbasedondensecorrespondenceandweightingwithsaliencevalues,andthe1.Introductionlowerpartshowssomepairsofimageswiththeirsaliencemaps.Personre-identificationhandlesp

7、edestrianmatchingandrankingacrossnon-overlappingcameraviews.Ithasmanyimportantapplicationsinvideosurveillancebysavingalotbyemployingsupervisedmodels,whichrequiretrainingofhumaneffortsonexhaustivelysearchingforapersondatawithidentitylabels.Also,mostofthemrequir

8、elabel-fromlargeamountsofvideosequences.However,thisisingnewtrainingdatawhencamerasettingschange,sincealsoaverychallengingtask.Asurveillancecameramaythecross-viewtransformsarediffe