FaceRecognitionAccuracy
AliceJ.O’Toole,FangJiang,HervéAbdi&P.JonathonPhillips∗
∗NationalInstituteofStandardsandTechnology
TheUniversityofTexasatDallas
Abstract.Recentworkindicatesthatstate-of-the-artfacerecognitionalgo-rithmscansurpasshumansmatchingidentityinpairsoffaceimagestakenunderdifferentilluminationconditions.Ithasbeendemonstratedfurtherthatfusingalgorithm-andhuman-derivedfacesimilarityestimatescutserrorratessubstantiallyovertheperformanceofthebestalgorithms.Hereweemployedapattern-basedclassificationproceduretofuseindividualhumansubjectsandalgorithmswiththegoalofdeterminingwhetherstrategydifferencesamonghumansarestrongenoughtosuggestparticularman-machinecombinations.Theresultsshowedthaterrorratesforthepairwiseman-machinefusionswerereducedanaverageof47percentwhencomparedtotheperformanceofthealgorithmsindividually.Theperformanceofthebestpairwisecombinationsofindividualhumansandalgorithmswasonlyslightlylessaccuratethanthecombinationofindividualhumanswithallsevenalgorithms.Thebalanceofmanandmachinecontributionstothepairwisefusionsvariedwidely,indicat-ingthataone-size-fits-allweightingofhumanandmachinefacerecognitionestimatesisnotappropriate....
1Introduction
Facerecognitionalgorithmshaveimproveddramaticallyoverthelastdecadeandarenowavailablecommerciallyforsecurityapplications.Themostcom-monapplicationsinvolvefaceverification,whereapresentedimageofaper-sonmustbecomparedtoastoredrepresentationandbeverifiedorrejectedasanidentitymatch.Toachievethistask,algorithmsproduceanestimateofthelikelihoodthattwoimagesareofthesameperson.Amatchcriterionisthensetformakingthedecision.
AccordingtoarecentUSGovernmentsponsoredtestofstate-of-the-artfacerecognitionalgorithms[1,2],currentalgorithmsareimpressivelyaccu-rateatthistaskwhentheimagesto-be-matchedaretakenundercontrolledilluminationconditions.Specifically,theFaceRecognitionGrandChallenge
⋆InG.Bebisetal.(Eds):AdvancesinVisualComputing.ISVC2006.LNCS4292.NewYork:
SpringerVerlag.pp.447-456,2006.
Dowloadedfromwww.utdallas.edu/∼herve
448
AliceJ.O’Toole,FangJiang,HervéAbdi&P.JonathonPhillips∗
(FRGC),conductedattheNationalInstituteofStandardsandTechnology(NIST)between2004and2006,tested17algorithmsonataskofmatchingfaceidentityinroughly128millionpairsofimagestakenundercontrolledil-lumination.Theresultsshowedanaverageverificationrateof.91atthe.001falseacceptancerate.
InananalogousFRGCtestofalgorithmsonthetaskofmatchingfaceidentityinimagestakenunderdifferentilluminationconditions,however,theperformanceofalgorithmswaslessimpressive.Onlysevenalgorithmsvolunteeredtoparticipateinthisexperiment.Thesesystemsscoredanaver-ageverificationrateofonly.42atthe.001falseacceptancerate.Thissuggeststhatcurrentfacerecognitionalgorithmsmaynotbereadyforapplicationen-vironmentsthathavenaturalvariationsinillumination.Thesekindsofen-vironmentsaretypicalinairportsandotherpublicplaceswheresomepartoftheilluminationcomesfromnaturallight,(e.g.,sunlightfilteredthroughwindows).
Howwellmustafacerecognitionalgorithmperformtobeusefulforase-curityapplication?Althoughhumanperformanceonfacerecognitionisof-tenconsideredthestandardtowhichalgorithmsshouldaspire,therearefewdirectcomparisonsbetweenhumansandalgorithms.Arecentexceptiontothisgeneralruleisastudycomparingtheperformanceofhumanswithalgo-rithmscompetingintheuncontrolledilluminationexperimentoftheFRGC[3].Inthatstudy,“easy”and“difficult”facepairsweresampledfromtheFRGCtestset,usingacontrolalgorithmbasedonprincipalcomponentsanalysis(PCA)ofthealignedandscaledimages.Humansubjectsratedthelikelihoodthatthepairsoffaceimageswereofthesameperson.ROCcurvescomputedforthehumansandforthealgorithmsrevealedthatthreealgorithms[4–6]performedmoreaccuratelythanhumansonthedifficultfacepairs.Nearlyallalgorithmsperformedmoreaccuratelythanhumansontheeasyfacepairs.Post-hocanalysesofthehumansubjectdatagavenoindicationthatsubjectattentionwanedtowardtheendoftheexperiment.Moreover,bothhumansandalgorithmsweremoreaccurateonfacepairsprescreenedbythePCAtobeeasythanonthefacepairsprescreenedtobedifficult.Theresultsindi-cate,therefore,thatalthoughalgorithmsappeartoperformpoorlyonthetask,theyarenonethelesscompetivewiththeperformanceofhumansub-jects.
Thecomparisonbetweenalgorithmsandhumansonthefacematchingtaskgivesanindicationofthequantitativerankingofalgorithmsbyaccuracy,butdoesnotofferanyinformationabouthowsimilarlyhumansandalgo-rithmsperformthetask.Togaininsightintothequalitativeaspectsoftheperformanceofhumansandalgorithms,andtoseeifperformancecouldbe
LectureNotesinComputerScience449
improved,afusionstudyofthealgorithmsandhumanswascarriedout[7].Ifalgorithmsandhumansemploydifferentrecognitionstrategies,itshouldbepossibletofusethetheirestimatesoffacesimilaritytoimproveperfor-mance.Thefusionwascarriedoutintwoparts.Inthefirstpart,thesimilarityestimatesofthesevenalgorithmsfromtheFRGCwerefusedwithpartialleastsquaresregression(PLS)[8,9]andusedtopredictthematchstatusofindivid-ualpairsoffaces.TherobustnessofthePLSwastestedwithcross-validation.Theresultsindicatedthatfusingthealgorithmscuttheerrorrateofthebest-performingalgorithmbyafactoroftwo[7].
Thesecondpartofthestudyexaminedwhetheralgorithmperformancecouldbeimprovedbyfusinghumansimilarityestimateswiththeestimatesofthesevenalgorithms.Specifically,humansimilarityestimatesforthedifficultfacepairs[3]wereaveragedacross49subjectsandfusedwiththealgoritmsestimatesusingPLS.Thisreducederrorratetonearperfectionandindicatedthathumanfacerecognitionstrategiesdiffersufficientlyfromalgorithmstomakeasubstantialcontributiontorecognitionperformancethroughfusion.
Althoughitisofgeneralinteresttoknowthathumansandalgorithmscanbefusedtoincreasefacematchingaccuracy,itisofmorepracticalvaluetoknowhowtocombineindividualalgorithmswithindividualhumans.Whichalgorithmscombinemostbeneficiallywithwhichhumans?Canageneralrulebeestablished,ordohumansandalgorithmsdiffersufficientlyinrecog-nitionstrategytosuggestthatindividualhuman-machinefusionsbedoneonacase-by-casebasis.Thepurposeofthepresentstudywastoexploretheben-efitsoffusingindividualhumanswithindidvidualalgorithms.Thisisofprac-ticalvaluegiventhatinmostreal-worldapplications,onealgorithmworksunderthesupervisionofasinglehumanoperator.Weassessedthesuitabilityofparticularman-machinecombinationsusingafusionapproach.
Weapproachedthisproblemintwoparts.First,wefusedthesimilarityscoresproducedbythesevenavailablealgorithmsandindividualhumans.Thisprovidesdataonthebest-casescenario,(i.e.wherehumanscanben-efitfromallavailablealgorithmexpertise).Thisfusionissimilartotheonedescribedpreviously[7],withtheexceptionthathumans,inthiscase,aretreatedasindividuals,ratherthanbeingrepresentedbytheirglobalaverage.Next,wefusedindividualhumansubjectswithindividualalgorithmstoas-sessperformanceofparticularman-machinehybrids.
2Methods
ThemethodsforthisstudymakeuseofhumansubjectandFRGCalgorithms’estimatesoffacepairsimilaritythatwerecollectedpreviously[3].Forcom-
450
AliceJ.O’Toole,FangJiang,HervéAbdi&P.JonathonPhillips∗
pletenessweprovideasketchofthemethodsusedtocollecttherelevantdatainthatstudyandthenproceedtodescribethefusionmethodsappliedinthepresentwork.2.1Stimuli
ThefaceimagesusedtotestalgorithmsinthefacematchingtaskweredrawnfromadatabasedevelopedfortheFRGC[1].Facepairsconsistedofatargetimage,takenundercontrolledilluminationconditions,andaprobeimage,takenunderuncontrolledilluminationconditions,(e.g.,inacorridor).Targetimageshadaresolutionof1704x2272pixelsandprobeimageshadaresolu-tionof2272x1704.AsamplefacepairappearsinFigure1.
Fig.1.Anexamplenon-matchpairwiththeprobeimageontheleftandthetargetprobeimageontheright
Tomakethetaskaschallengingaspossible,wesampledtheimagesfromahomogenousfacepopulationthatincludedonlymaleandfemaleCauca-siansintheirtwentiesandthirties.Eachfacepairwasmatchedbysex.Com-binedtheseconstraintseliminatedthepossibilitythathumanscouldbaseidentityjudgmentsonsurfacefacialcharacteristicsassociatedwithsex,race,orage.
Inthepresentstudy,weusedthe120difficultpairsoffaces(60maleand60female)presentedtosubjectsinthepreviousstudy[3].Halfofthefacepairswereofmatchedidentity,(i.e.,thetargetandprobeimageswereofthesameperson),andhalfwerenon-matchpairs(i.e.,thetargetandprobeim-ageswereofdifferentpeople).Asnoted,facepairswereprescreenedbyaPCAofthealignedandscaledimagesbeforesamplingforthehumanexperi-ment.Difficultmatchpairshadsimilarityscoresthatwerelessthantwostan-
LectureNotesinComputerScience451
darddeviationsbelowthematchmean,(i.e.,twodissimilarpicturesofthesameperson).Difficultnon-matchpairshadsimilarityscoresgreaterthantwostandarddeviationsabovethenon-matchmean,(i.e.,similarimagesoftwodifferentpeople).
2.2Humanestimatesoffacesimilarity
Humansjudgedthesimilarityofthe120facepairsbyratingeachpairasfol-lows,“1.)suretheyarethesameperson;2.)thinktheyarethesameperson;3.)don’tknow;4.)thinktheyarenotthesameperson;5.)suretheyarenotthesameperson.”Theratingsof49subjectsonthe120facepairsservedasthehumaninputdataforthefusionanalysis.2.3Algorithmestimatesoffacesimilarity
ToparticipateintheFRGCuncontrolledilluminationexperiment,algorithmdeveloperswereaskedtocomputeamatrixcontainingsimilarityscoresbe-tweenallpossiblepairsof16,028targetimagesand8,014probeimages.Theresultingmatrix,therefore,contained128,448,392similarityscores,eachrep-resentingthelikelihoodthatatargetandprobeimagesinthepairwereofthesameperson.ThesematriceswerescoredbyNISTandcompleteperfor-manceresultsforthealgorithmsareavailableelsewhere[1,2].Forpresentpurposes,thesimilarityscoresforthe120difficultfacepairspresentedtosubjectswereextractedfromeachofsevenparticipatingalgorithms’similar-itymatrices.Thesescoresservedasthealgorithms’inputtothefusionanaly-sis.
2.4PLSfusion
Fusionwasperformedbypartialleastsquares(PLS)regression,astatisticaltechniquethatgeneralizesandcombinesfeaturesfromprincipalcomponentanalysisandmultipleregression[8,9].Thetechniqueisusedtopredictasetofdependentvariablesfromasetofindependentvariables(predictors).Thoughlessknowninthepatternrecognitionliterature,PLSiswidelyusedinchemometrics,sensoryevaluation,andneuroimagingdataanalysis,(cf.[9]).
Thepredictorsforthepresentstudywerethehuman-andalgorithm-generatedsimilarityscoresforthe120pairsoffaceimages.Wedefinethismorespecificallyinthecontextofparticularfusionanalyses.Thedependentvariablewasthematchstatusofthefacepair(i.e.,sameversusdifferentper-son).withmatchpairsassignedavalueof1andnon-matchpairsavalue
452
AliceJ.O’Toole,FangJiang,HervéAbdi&P.JonathonPhillips∗
of−1.PLSregressiongivesasetoforthogonalfactors,sometimescalledla-tentvectors,fromthecovariancematrixofpredictorsanddependentvari-ables.Thesecanbeusedtopredictthedependentvariable(s),byappropri-atelyweightingthepredictors.ThissetofweightsiscalledBplsinthePLS-regressionliterature[8,9].
ThepredictivepowerofaPLSsolutionisassessedgenerallywithcross-validationtechniquessuchasabootstraporjackknifeprocedure.Allfactors,oronlyasubsetofthem,canbeusedtocomputethepredictionofthede-pendentvariable(s),whichareobtainedasaweightedcombinationoftheoriginalpredictorsgivenbyBpls.Thelargerthenumberoffactorskept,thebetterthepredictionofthe“learningset”but,ingeneral,asmallernumberoffactorsisoptimalforrobustprediction(i.e.,fortestsetpredictions).
3Results
3.1BaselineHumanPerformance
Wefirstassessedthebaselineperformanceofindividualhumansinawaythatiscomparablewiththeoutputofthefusiontest.Accuracyinthefusiontestisgivenastheerrorrateformatchstatusclassificationsdeterminedinajack-knifeprocedure.Asnoted,humansratedeachpairoffacesona5-pointscalevaryingfrom“surethesameperson”to“suredifferentpeople”.Giventhatoneofthepossiblehumanresponseswas“don’tknow”,wecom-putedthenumberoferrorsintwowaysby:a.)assigningthe“don’tknow”toamatchresponse;andb.)assigningthe“don’tknow”responsetoanon-matchresponse.Wecomputedthenumberoferrorsforeachsubjectinbothwaysandaveragedthesetwovalues.Acrossallsubjects,thehumanerrorrateaveraged.141(seecolumnoneofTable1),indicatinggood,butnotperfectperformance.
Table1.Theerrorrateinformationforindividualhumanperformanceappearsinthefirstcolumn.Analogousinformationappearsinthesecondcolumnforthefusionofindividualhumanscombinedwiththesevenalgorithms.Thedataforthebest-pairedindividualhuman-algorithmfusionsappearinthethirdcolumn
Errorrateaverageminimummaximum
HumanHuman+7Algorithms.141.041.258
.057.033.108
Human+AlgorithmBest-pairedFusions
.078.033.117
LectureNotesinComputerScience453
3.2FusingIndividualHumanswithAllAlgorithms
Next,wefusedeachindividualwithallsevenalgorithmsusingPLS.Thepur-poseofthisanalysiswastoassesstherelativeimportanceofalgorithmswhencombinedwithindividualhumans.Wealsousetheperformanceofthissimu-lationas“bestcase”controlfortheindividualhuman-algorithmfusionsthatfollow.
ApredictormatrixforthePLSwascreatedforeachofthe49subjectsasfollows.Estimatesoffacesimilarityforthe120facepairsfromthehumansubjectwerecombinedinacolumn-wisematrixwiththeestimatestakenfromthesevenalgorithmsforthesamefacepairs.Ajack-knifeprocedureoperatedbydeletingeachofthefacepairsinturnandcomputingthePLSontheremaining119pairs.Accuracywasmeasuredasthenumberofcor-rectclassificationsforthedeleted,“left-out”facepairs.Inallcases,jack-knifeprocedurestestedtheaccuracyofhumanalgorithmfusionwith1through5factorsPLSsolutions.Theoptimalnumberoffactorsforthe49subjectsvar-iedfrom1to5,withamedianof1,andameanof1.36.Foreachsubject,onlythesolutionthatyieldedhighestaccuracyisreported.
Fusingindividualhumanswiththesevenalgorithmscutthehumanerrorratebymorethanafactoroftwo(cf.,Table1).Forreference,theerrorratesfortheindividualalgorithmsfrom[7]appearinfirstcolumnofTable2.Ascanbeseen,theaverageerrorrateachievedbyfusingindividualsubjectswiththecombinationofexpertisefoundinthesevenalgorithmsislessthanhalfofthatachievedbythebestalgorithm.
Tointerprettheroleofindividualalgorithmsandhumansintbefusion,welookedatthePLS-derivedweightsforcombiningsimilarityestimatesinthePLS.TheseweightsemergefromthePLSasarecipeforoptimallycom-biningalgorithmandhumansimilarityestimatestopredictmatchstatus.Av-eragedacrossthe49subject-basedfusions,thestrongestcontributortothefusionwasthealgorithmfromViisageCorporation[6].ThiswasfollowedbythealgorithmfromtheNewJerseyInstituteofTechnology(NJIT)[4].Anony-mousalgorithmsBandDalsoplayedaroleinthefusion.Theweightsforhu-manswerewellbelowthesealgorithmweightsindicatingthattheroleofhu-mansinthefusionwasminor.Inpreviouswork[7],wefusedtheaveragehu-manwiththesevenalgorithmsandfoundthatthehumanroleinimprovingperformanceinthefusionwasmoresubstantial.Herewefoundthatfusingindividualsubjectperformancewiththesevenalgorithmsandaveragingtheweightsafterwardsrevealedamuchsmallerroleforhumans.Thissuggeststhatindividualhumanperformancemayvarystrategicallyandthatpartic-ularperson-tailoredmixturesofindividualsandalgorithmsmayprovideforbetterbalanceinhuman-machinefusions.Thisfindingmotivatesthenext
454
AliceJ.O’Toole,FangJiang,HervéAbdi&P.JonathonPhillips∗
Table2.Resultscompilationforhuman-algorithmfusions.Thefirstcolumncontainserrorratesobtainedpreviously[7]foreachalgorithmoperatingalone.ThehumanerrorratefromTable1appearsatthetopofthecolumnforcomparison.Thesecondcolumnshowstheaver-ageerrorratesachievedbyfusingindividualhumanswitheachofthesevenalgorithms.Thepercentagereductioninerrorratefromtheindividualalgorithmstohuman-algorithmpairfusionappearsincolumn3.
SourcehumanNJITViisageCMU
AlgorithmAAlgorithmBAlgorithmCAlgorithmD
individualerrorrate[7]
.14.12.20.14.37.23.25.26
pairedfusion%errorrateerrorratereduction
–.08.12.09.13.11.12.13
33%40%36%65%52%52%50%
analysisinwhichwefusedallpossiblepairsofindividualhumansandindi-vidualalgorithms.
3.3FusingIndividualHumansandIndividualAlgorithms
Thefusionwascarriedoutasdescribedpreviously,butthistimecombiningeachofthe49subjectsinidviduallywitheachofthe7algorithms.Bydefini-tion,only1and2-factorsolutionsarepossible,andsowecarriedoutapre-testofallhuman-modelcombinationsusingPLStodeterminetheoptimalnumberoffactorsforeachpairing.Inwhatfollows,onlythebetterofthesetwosolutionsisanalyzed.
Theaverageperformanceforthebestpair-wisefusionbetweenindivid-ualhumansubjectsandalgorithmsappearsincolumn3ofTable1.Perfor-manceforthisfusionisreducedonlyslightlyfromtheperformancewefoundwhencombiningindividualswithallsevenalgorithms.Thisindicatesthattheperformanceofpair-wisefusionscancomeclosetothatachievedbyfusinghumanswithallavailablealgorithmexperitise.Forreference,theaverageper-formanceofpossiblepairsofhuman-algorithmfusionswas0.111,indicatingthatthebestpair-wisefusionsweresubstantiallybetterthantheoverallaver-ageofpaired-fusions.Thissuggeststhathuman-machinefusionsarenotallequallybeneficial.
Wenotealsothatinpreviouswork[7],fusingthe7algorithms(withouthumans)producedanerrorrateof.059.Thisiscomparabletotheerrorrate
LectureNotesinComputerScience455
ofindividualhumansfusedwiththesevenalgorithms,butissubstantiallylargerthanthenear-zeroerrorrateachievedwhenfusingtheaveragehumanwiththe7algorithms[7].
Overall,theNJITalgorithm[4]producedthebestperformancewhenpairedwithindividualhumansubjects.Fortyofthe49subjectsproducedtheirbestfusedperformancewiththisalgorithm.Anadditional8subjectspairedbestwithCarnegieMellonUniversity’salgorithm[5]andonepersonpairedbestwithAlgorithmB.Thatsaid,whencomparingtheperformanceofthealgo-rithmsindividually(column1ofTable2)withtheperformanceofthepairedhuman-algorithmfusions(column3ofTable2,theerrorratereductionsareremarkable(column4ofTable2showspercentageerrorratereductionsforthesevenalgorithms).Onaverage,theerrorrateisreducedby47percent.Clearly,muchofthisimprovementcomesfromthereductionoferrorrateforlesserperformingalgorithms.Itbearsnoting,however,thatevenwiththelargeerrorratereductions,theaverageperformanceofahumanaloneisonlyslightlyworsethanthepairedfusionsforthelesserperformingalgorithms.Giventhatthepairedfusionsareworthwhileonlywhentheybettertheer-rorratesofthealgorithmorhumanoperatingalone,thepairwisefusionsofCMU[6]andNJIT[4]arethemostpromisingcandidatesforfusion.
Finally,thebalanceofmanandmachineinputinthesefusionscanbeas-sessedbylookingatthePLS-derivedweightsforthepairwisefusions.ThesearedisplayedinTable3,withthePLS-derivedweightsforthehumanandal-gorithmscoresinthefirstandsecondcolumns,respectively.Theproportionofhumancontributiontothesefusionsisgivenincolumn3.Thehumancon-tributionvariesfromaminimumof.0439forViisage[6]toalmostcompletedominanceforanonymousalgorithmsAandC.Notably,theman-machinebalancevarieswidelywiththealgorithm.
4Discussion
Inmostsecurityapplications,theuseoffacerecognitionalgorithmsisun-derthesupervisionofahumanoperator.Previousstudiesindicatethattheperformanceofalgorithmscancompetewithhumansinsomecases[3].Theaccuracyofindividualhumansandindividualalgorithms,however,canbequitevariable.Inthesecases,thequestionofwhetheralgorithmsorhumansperformmoreaccuratelymaybelessimportantthanunderstandinghowpar-ticularman-machinecombinationsperform.
Inthisstudy,wefusedindividualalgorithmsandindividualhumansandshowthatthesepairwisefusionsperformedsubstantiallybetterthantheal-gorithmsoperatingalone.Thissuggeststhathumanscancontribute,through
456
AliceJ.O’Toole,FangJiang,HervéAbdi&P.JonathonPhillips∗
Table3.Thebalanceofhuman-algorithmcontributionsinthepairwisefusions.Theweightsforhumansandalgorithmsappearinthefirstandsecondcolumns,respectively.Thethirdcolumngivestheproportionofhumancontributiontotheoverallfusion.Thisproportionis
Humancomputedas
Human+AlgorithmSourceNJITViisageCMU
AlgorithmAAlgorithmBAlgorithmCAlgorithmD
humanalgorithmproportionweightweighthuman−0.2683−0.3268−0.2992−0.4073−0.3283−0.3422−0.4052
−2.4902−7.1175−0.0900−0.0025−2.4525−0.0009−2.3027
0.09730.04390.76890.99400.11800.99740.1496
fusion,tobetterfacerecognitionperformance.Onecaveat,however,isthatthesepairwisefusionsmustcompete,notonlywiththeperformanceofthealgorithmalone,butwiththeperformanceofthehumanalone.Fromthisperspective,pairwisefusionscompetewithbothhumansandalgorithmsonlyforthebestperformingalgorithms.
Finally,weshowthatitisimportanttocombinehumanandalgorithm-generatedresponsesinquantitativelyprecisewaystoimproveperformanceoptimally.Thestrongvariabilityinman-machinebalancewefoundacrossthe7algorithmsillustratestheimportanceofconsideringalgorithmsandhu-mansasindividuals.
References
1.Phillips,P.,Flynn,P.,Scruggs,T.,Bowyer,K.,Chang,K.,Hoffman,J.,Marques,J.,Min,J.,Worek,W.:Overviewofthefacerecognitiongrandchalleng.In:ProceedingsoftheIEEEComputerVisionandPatternRecognition.(2005)947–954
2.Phillips,P.,Flynn,P.,Scruggs,T.,Bowyer,K.,Worek,W.:Preliminaryfacerecognitiongrandchallengeresults.In:ProceedingsoftheSeventhInternationalConferenceonAutomaticFaceRecognition.(2006)15–24
3.O’Toole,A.,Phillips,P.,Jiang,F.,Ayyad,J.,Penard,N.,Abdi,H.:Facerecognitionalgorithmssurpasshumans.submitted(2006)
4.Liu,C.:Capitalizeondimensionalityincreasingtechniquesforimprovingfacerecognitiongrandchallengeperformance.IEEE:TransactionsonPatternAnalysisandMachineIntelli-gence(2006)725–737
5.Xie,C.,Savvides,M.,Kumar,V.:Kernelcorrelationfilterbasedredundantclass-dependencefeatureanalysis(kcfa)onfrgc2.0data.IEEEInternationalWorkshoponAnalysisandMod-elingFacesandGestures1(2005)32–43
LectureNotesinComputerScience457
6.Husken,M.,Brauckmann,B.,Gehlen,S.,vonderMalsburg,C.:Strategiesandbenefitsoffusionof2dand3dfacerecognition.In1,ed.:ProceedingsoftheIEEEWorkshoponFaceRecognitionGrandChallengeExperiments.ComputerSocietyDigitalLibrary.Volume3.,IEEEPress(2005)174
7.O’Toole,A.,Abdi,H.,F.,J.,Phillips,P.:Fusingfacerecognitionalgorithmsandhumans.submitted(2006)
8.Abdi,H.:Partialleastsquaresregression.InBeck,M.,A.,B.,Futing,T.,eds.:EncyclopediaforResearchMethodsintheSocialSciences,ThousandOaks,CA,Sage(2003)792–7959.McIntosh,A.,Lobaugh,N.:Partialleastsquaresanalysisofneuroimagingdata:applica-tionsandadvances.Neuroimage23(2004)250–263
因篇幅问题不能全部显示，请点此查看更多更全内容