2010-2016精選論文2014-deepface-closing-the-gap-to-human-level-performance

1、DeepFace:ClosingtheGaptoHumanLevelPerfmanceinFaceVerificationYanivTaigmanMingYangMarc’AurelioRanzatoFacebookAIResearchMenloParkCAUSAyanivmingyangranzato@LiWolfTelAvivUniversityTelAvivstractInmodernfacerecognitiontheconve

2、ntionalpipelineconsistsoffourstages:detect?align?represent?classify.Werevisitboththealignmentsteptherepresentationstepbyemployingexplicit3Dfacemodelingindertoapplyapiecewiseaffinetransfmationderiveafacerepresentationfrom

3、aninelayerdeepneuralwk.Thisdeepwkinvolvesmethan120millionparametersusingseverallocallyconnectedlayerswithoutweightsharingratherthanthestardconvolutionallayers.Thuswetraineditonthelargestfacialdatasettodateanidentitylabel

4、eddatasetoffourmillionfacialimagesbelongingtomethan4000identities.Thelearnedrepresentationscouplingtheaccuratemodelbasedalignmentwiththelargefacialdatabasegeneralizeremarkablywelltofacesinunconstrainedenvironmentsevenwit

5、hasimpleclassifier.Ourmethodreachesanaccuracyof97.35%ontheLabeledFacesintheWild(LFW)datasetreducingtheerrofthecurrentstateoftheartbymethan27%closelyapproachinghumanlevelperfmance.1.IntroductionFacerecognitioninunconstrai

6、nedimagesisatthefefrontofthealgithmicperceptionrevolution.Thesocialculturalimplicationsoffacerecognitiontechnologiesarefarreachingyetthecurrentperfmancegapinthisdomainbetweenmachinesthehumanvisualsystemservesasabufferfro

7、mhavingtodealwiththeseimplications.Wepresentasystem(DeepFace)thathasclosedthemajityoftheremaininggapinthemostpopularbenchmarkinunconstrainedfacerecognitionisnowatthebrinkofhumanlevelaccuracy.Itistrainedonalargedatasetoff

8、acesacquiredfromapopulationvastlydifferentthantheoneusedtoconstructtheevaluationbenchmarksitisabletooutperfmexistingsystemswithonlyveryminimaladaptation.Meoverthesystemproducesanextremelycompactfacerepresentationinsheerc

9、ontrasttotheshifttowardtensofthoussofappearancefeaturesinotherrecentsystems[572].Theproposedsystemdiffersfromthemajityofcontributionsinthefieldinthatitusesthedeeplearning(DL)framewk[321]inlieuofwellengineeredfeatures.DLi

10、sespeciallysuitablefdealingwithlargetrainingsetswithmanyrecentsuccessesindiversedomainssuchasvisionspeechlanguagemodeling.Specificallywithfacesthesuccessofthelearnedincapturingfacialappearanceinarobustmannerishighlydepen

11、dentonaveryrapid3Dalignmentstep.Thewkarchitectureisbasedontheassumptionthatoncethealignmentiscompletedthelocationofeachfacialregionisfixedatthepixellevel.ItistherefepossibletolearnfromtherawpixelRGBvalueswithoutanyneedto

12、applyseverallayersofconvolutionsasisdoneinmanyotherwks[1921].Insummarywemakethefollowingcontributions:(i)Thedevelopmentofaneffectivedeepneural(DNN)architecturelearningmethodthatleverageaverylargelabeleddatasetoffacesinde

13、rtoobtainafacerepresentationthatgeneralizeswelltootherdatasets(ii)Aneffectivefacialalignmentsystembasedonexplicit3Dmodelingoffaces(iii)Advancethestateoftheartsignificantlyin(1)theLabeledFacesintheWildbenchmark(LFW)[18]re

14、achingnearhumanperfmance(2)theYouTubeFacesdataset(YTF)[30]decreasingtheerrratetherebymethan50%.1.1.RelatedWkBigdatadeeplearningInrecentyearsalargenumberofphotoshavebeencrawledbysearchenginesuploadedtosocialwkswhichinclud

15、eavarietyofunconstrainedmaterialsuchasobjectsfacesscenes.Thislargevolumeofdatatheincreaseincomputationalresourceshaveenabledtheuseofmepowerfulstatisticalmodels.Thesemodelshavedrasticallyimprovedtherobustnessofvisionsyste

16、mstoseveralimptantvariationssuchasnonrigiddefmationsclutterocclusionilluminationallproblemsthatareattheceofmanycomputervisionapplications.Whileconventionalmachine1pointdetectbutapplyitinseveraliterationstorefineitsoutput

17、.AteachiterationfiducialpointsareextractedbyaSupptVectRegress(SVR)trainedtopredictpointconfigurationsfromanimagede.OurimagedeisbasedonLBPHistograms[1]butotherfeaturescanalsobeconsidered.Bytransfmingtheimageusingtheinduce

18、dsimilaritymatrixTtoanewimagewecanrunthefiducialdetectagainonanewfeaturespacerefinethelocalization.2DAlignmentWestartouralignmentprocessbydetecting6fiducialpointsinsidethedetectioncropcenteredatthecenteroftheeyestipofthe

19、nosemouthlocationsasillustratedinFig.1(a).TheyareusedtoapproximatelyscalerotatetranslatetheimageintosixanchlocationsbyfittingTi2d:=(siRiti)where:xjanch:=si[Ri|ti]?xjsourcefpointsj=1..6iterateonthenewwarpedimageuntilthere

20、isnosubstantialchangeeventuallycomposingthefinal2Dsimilaritytransfmation:T2d:=T12d?...?Tk2d.Thisaggregatedtransfmationgeneratesa2DalignedcropasshowninFig.1(b).ThisalignmentmethodissimilartotheoneemployedinLFWawhichhasbee

21、nusedfrequentlytoboostrecognitionaccuracy.Howeversimilaritytransfmationfailstocompensatefoutofplanerotationwhichisparticularlyimptantinunconstrainedconditions.3DAlignmentIndertoalignfacesundergoingoutofplanerotationsweus

22、eageneric3Dshapemodelregistera3Daffinecamerawhichareusedtowarpthe2Dalignedcroptotheimageplaneofthe3Dshape.Thisgeneratesthe3DalignedversionofthecropasillustratedinFig.1(g).Thisisachievedbylocalizingadditional67fiducialpoi

23、ntsx2dinthe2Dalignedcrop(seeFig.1(c))usingasecondSVR.Asa3Dgenericshapemodelwesimplytaketheaverageofthe3DscansfromtheUSFHumanIDdatabasewhichwerepostprocessedtoberepresentedasalignedverticesvi=(xiyizi)ni=1.Wemanuallyplace6

24、7anchpointsonthe3Dshapeinthiswayachievefullcrespondencebetweenthe67detectedfiducialpointstheir3Dreferences.Anaffine3Dto2DcameraPisthenfittedusingthegeneralizedleastsquaressolutiontothelinearsystemx2d=X3d?Pwithaknowncovar

25、iancematrixΣthatis?Pthatminimizesthefollowingloss:loss(?P)=rTΣ?1rwherer=(x2d?X3d?P)istheresidualvectX3disa(67?2)8matrixcomposedbystackingthe(28)matrices[x?3d(i)1?0?0x?3d(i)1]with?0denotingarowvectoffourzerosfeachreferenc

26、efiducialpointx3d(i).TheaffinecameraPofsize24isrepresentedbythevectof8unknowns?P.ThelosscanbeminimizedusingtheCholeskydecompositionofΣthattransfmstheproblemintodinaryleastsquares.Sincefexampledetectedpointsonthecontourof

27、thefacetendtobemenoisyastheirestimatedlocationislargelyinfluencedbythedepthwithrespecttothecameraangleweusea(67?2)(67?2)covariancematrixΣgivenbytheestimatedcovariancesofthefiducialpointerrs.FrontalizationSincefullperspec

28、tiveprojectionsnonrigiddefmationsarenotmodeledthefittedcameraPisonlyanapproximation.Indertoreducethecruptionofsuchimptantidentitybearingfactstothefinalwarpingweaddthecrespondingresidualsinrtothexycomponentsofeachreferenc

29、efiducialpointx3dwedenotethisas?x3d.Sucharelaxationisplausiblefthepurposeofwarpingthe2Dimagewithsmallerdisttionstotheidentity.Withoutitfaceswouldhavebeenwarpedintothesameshapein3Dlosingimptantdiscriminativefacts.Finallyt

30、hefrontalizationisachievedbyapiecewiseaffinetransfmationTfromx2d(source)to?x3d(target)directedbytheDelaunaytriangulationderivedfromthe67fiducialpoints1.Alsoinvisibletrianglesw.r.t.tocameraPcanbereplacedusingimageblending

31、withtheirsymmetricalcounterparts.3.RepresentationInrecentyearsthecomputervisionliteraturehasattractedmanyresearchefftsindeengineering.Suchdeswhenappliedtofacerecognitionmostlyusethesameoperattoalllocationsinthefacialimag

32、e.Recentlyasmedatahasbecomeavailablelearningbasedmethodshavestartedtooutperfmengineeredfeaturesbecausetheycandiscoveroptimizefeaturesfthespecifictaskath[19].Herewelearnagenericrepresentationoffacialimagesthroughalargedee

33、pwk.DNNArchitectureTrainingWetrainourDNNonamulticlassfacerecognitiontasknamelytoclassifytheidentityofafaceimage.TheoverallarchitectureisshowninFig.2.A3Daligned3channels(RGB)faceimageofsize152by152pixelsisgiventoaconvolut

34、ionallayer(C1)with32filtersofsize11x11x3(wedenotethisby32x11x11x3@152x152).Theresulting32featuremapsarethenfedtoamaxpoolinglayer(M2)whichtakesthemaxover3x3spatialneighbhoodswithastrideof2separatelyfeachchannel.Thisisfoll

35、owedbyanotherconvolutionallayer(C3)thathas16filtersofsize9x9x16.Thepurposeofthesethreelayersistoextractlowlevelfeatureslikesimpleedgestexture.Maxpoolinglayersmaketheoutputofconvolutionwksmerobusttolocaltranslations.Whena

36、ppliedtoalignedfacialimagestheymakethewkmerobusttosmallregistrationerrs.Howeverseverallevelsofpoolingwouldcausethewktoloseinfmationabouttheprecisepositionofdetailedfacialstructuremicrotextures.Henceweapplymaxpoolingonlyt