1 Three-dimensional caricatures of human heads Dis

1Three-dimensionalcaricaturesofhumanheads:DistinctivenessandtheperceptionoffacialageAliceJ.O’TooleTheUniversityofTexasatDallasThomasVetterMaxPlanckInstituteforBiologicalCyberneticsHaraldVolzMaxPlanckInstituteforBiologicalCyberneticsElizabethM.SalterTheUniversityofTexasatDallasMarch16,1997Facialageingandcaricatures2AbstractWeappliedastandardfacialcaricaturingalgorithmtoathree-dimensionalrep-resentationofhumanheads.Thisalgorithmsometimesproducedheadsthatap-peared\caricatured.Morecommonly,however,exaggeratingthedistinctivethree-dimensionalinformationinafaceseemedtoproduceanincreaseintheapparentageoftheface|bothatalocallevel,byexaggeratingsmallfacialcreasesintowrin-kles,andatamoregloballevelviachangesthatseemedtomaketheunderlyingstructureoftheskullmoreevident.Concomitantly,de-emphasisofthedistinctivethree-dimensionalinformationinafacemadeitappearrelativelyyoungerthantheveridicalandcaricaturedfaces.Moreformally,faceagejudgementsmadebyhu-manobserverswereorderedaccordingtothelevelofcaricature,withanti-caricaturesjudgedyoungerthanveridicalfaces,andveridicalfacesjudgedyoungerthancarica-turedfaces.Wediscusstheseresultsintermsoftheimportanceofthenatureofthefeaturesmademoredistinctbyacaricaturingalgorithmandthenatureofhumanrepresentation(s)offaces.Facialageingandcaricatures31IntroductionFacialcaricaturesareusedcommonlybyartiststoaccentuateorexaggeratethedistinctiveinformationinindividualfaces.Theautomatedproductionofcaricatureshasbeenpossibleformanyyearsduetotherelativesimplicityofthealgorithmsneededtomakethem(e.g.,Brennan,1985).Typically,suchalgorithmsoperateasfollows.First,ameasureoftheaveragevalueofasetof\featuresacrossalargenumberoffacesiscomputed.Thesefeaturesaredened,usually,asasetoffaciallandmarklocations(e.g.,cornersoftheeyeandotherpointsthatarereasonablyeasytolocalize/matchonallfaces)1inthetwo-dimensionalimage.Next,tocreateacaricatureofanindividualface,ameasureofthedeviationofthefacefromtheaveragetwo-dimensionalcongurationiscomputed.Finally,\distinctiveorunusualfeaturesofthefaceareexaggeratedtoproducethecaricature.Thisgenericalgorithmhasbeenappliedbothtolinedrawingsoffaces(e.g.,Bren-nan,1985;Rhodes,Brennan,&Carey,1987;Benson&Perrett,1994)andtopho-tographicqualityimages(e.g.,Benson&Perrett,1991);bothrepresentationsyieldperceptuallycompellingcaricatures.Moreformally,thereisevidencefrompsycholog-icalstudiesthatcaricaturesoffacesarerecognizedmorequicklyandaccuratelythanveridicalimagesoffaces(Benson&Perrett,1994;Mauro&Kubovy,1992;Stevenage,1995)andfurther,areratedas\betterlikenessesofindividualsthanveridicalimages(Benson&Perrett,1994).1ThoughseealsoBurt&Perrett,1995foranapplicationofthistechniquetoimageintensities.Facialageingandcaricatures4Facedistinctivenesseectshaveguredprominentlyinmanytheoreticalaccountsofhumanfaceprocessing(e.g.,Bruce&Young,1986;Goldstein&Chance,1980;Morton&Johnson,1991;Valentine,1991).Severalstudieshavedemonstratedthatfacesratedasdistinctbyobserversarebetterrecognizedthanfacesratedastypical(e.g.,Light,Kayra-Stuart,&Hollander,1979).Computer-generatedcaricatureshavebeenofinteresttopsychologistsbecausetheyprovideadirectmethodfortestingtheroleofdistinctivenessinfaceperceptionandrecognitiontasks(seeStevenage,1995;O’Toole&Edelman,1996).Thisdirectcontrolofdistinctivenessisanimportantadvantageofcaricaturestudiesbycomparsiontoratingstudies.Studiesinwhichthedistinctivenessofindividualfaceshasbeenmanipulateddirectlyviacaricaturinglendmoresupporttotheclaimthatfacedistinctivenessisanimportantfactorforhumanfaceperceptionandrecognition(PerrettandBenson,1991;1994;Rhodesetal.1987,Stevenage,1995).Theprimarydierencebetweenacaricatureapproachandassessingtherelation-shipbetweenratingandperformancedata(e.g.,Lightetal.,1979)istheexplicitnessofthedenitionof\distinctivenessrequiredinthecaricatureapproach.Specically,toapplyacaricaturealgorithmtofaces,onemustrstoperationallydenethein-formationinfacestobeexaggeratedorcaricatured.Thisisdone,typically,intermsofasetoflandmarkfacial\features,e.g.,cornersoftheeyes,tipofnose,thatcanbelocatedintheimageandaltered.Asnoted,althoughcaricaturealgorithmshavebeenappliedtodierentqualitiesofimagedata(linedrawingsversusphotographs),Facialageingandcaricatures5theyhavenotbeenappliedgenerallytofeaturesotherthanthosebasedonthetwo-dimensionalconguralstructureoffaces(thoughwediscussanexception,Burt&Perrett,1995,inSection4).Thepossiblilityofexploringtheperceptionofcaricaturesmadefromaninherentlythree-dimensionalrepresentationoffacesisinterestingforseveralreasons.First,thequestionoftheextenttowhichhumanrepresentationsofobjectsandfacesencodetwo-versusthree-dimensional\featureshasbeenverymuch-debatedinthepsy-chologyliteraturesincetheimportantpapersofBiederman(1987)andBulthoandEdelman(1992).Second,theserepresentationquestionsarenowbeinginvestigatedactivelybyneuroscientistsinterestedintheneurophysiologicalsubstratesofobjectandfacerecognition(Logothetis,Pauls,Poggio,1995;Perrett,Hietanen,Oram,&Benson,1992).Acomputationallybasedapproachfordeningandalteringfacedis-tinctiveness,incombina