Table-top computed lighting for practical digital photography

TabletopComputedLightingforPracticalDigitalPhotography

AnkitMohan,ReynoldBailey,JonathanWaite,JackTumblin,Member,IEEE,CindyGrimm,Member,IEEE,andBobbyBodenheimer,SeniorMember,IEEE

Abstract—Weapplysimplifiedimage-basedlightingmethodstoreducetheequipment,cost,time,andspecializedskillsrequiredforhigh-qualityphotographiclightingofdesktop-sizedstaticobjectssuchasmuseumartifacts.Weplacetheobjectandacomputer-steeredmoving-headspotlightinsideasimplefoam-coreenclosureanduseacameratorecordphotosasthelightscanstheboxinterior.Optimization,guidedbyinteractiveusersketching,selectsasmallsetofthesephotoswhoseweightedsumbestmatchestheuser-definedtargetsketch.Unlikepreviousimage-basedrelightingefforts,ourmethodrequiresonlyasinglearealightsource,yetitcanachievehigh-resolutionlightpositioningtoavoidmultiplesharpshadows.Areducedversionusesonlyahandheldlightandmaybesuitableforbattery-poweredfieldphotographyequipmentthatfitsintoabackpack.

IndexTerms—Image-basedlighting,enclosurelighting,handheldlighting,controllablelighting,digitalphotography.

Ç

1

INTRODUCTION

ODERNdigital

camerashavemadepicturetakingmuch

easierandmoreinteractive.However,lightingasceneforgoodphotographyisstilldifficult,andpracticalmethodstoachievegoodlightinghavechangedverylittleoverseveraldecades.Forthebestpossiblelighting,professionalphotographerschoosefromalargecollectionoflightsourcesofdifferentpower,shape,andsize,withdiversemountings,clamps,andstandsthatpermiteasyplacement,heightadjustment,andaiming.Theymayalsoattachattenuators,occluders,diffusers,filters,andreflec-torstosoftenorsharpen,brightenordarken,redirect,andaddcolortinttotheselightsources.Oncecomplete,allofthedesiredlightingeffectsarevisiblethroughtheview-finderallatonce,readyforcaptureintoasinglepicture.Unfortunately,goodstudiolightingcanbequitedifficultandtimeconsumingtoachieve.Athoroughlyequippedstudioisexpensive,anditmayencourageendlessrevisionandexperimenting:Whatwilllooknice?Whatmightlookalittlebitbetter?Howmightwebrightenonesceneregionaswedarkenanother?Afterafewhoursofwranglinghotlights,mostnonexpertsfindthatgoodstudiolightingcanbesurprisinglytediousandfrustrating.Mostpeoplecanspecifythelightingtheywantinscreenspace(forexample,

.A.MohanandJ.TumblinarewiththeDepartmentofElectricalEngineeringandComputerScience,NorthwesternUniversity,3.320FordDesignCenter,2133SheridanRoad,Evanston,IL60208.E-mail:{ankit,jet}@cs.northwestern.edu.

.R.BaileyandC.GrimmarewiththeDepartmentofComputerScienceandEngineering,WashingtonUniversityinSt.Louis,CampusBox1045,OneBrookingsDrive,St.Louis,MO63130.E-mail:{rjb1,cmg}@cse.wustl.edu..J.WaiteandB.BodenheimerarewiththeDepartmentofElectricalEngineeringandComputerScience,VanderbiltUniversity,VUStationB#351679,2301VanderbiltPlace,Nashville,TN37235.E-mail:{jon.waite,bobby.bodenheimer}@vanderbilt.edu.Manuscriptreceived15Mar.2006;revised29Aug.2006;accepted17Oct.2006;publishedonline8Jan.2007.

Forinformationonobtainingreprintsofthisarticle,pleasesende-mailto:tvcg@computer.org,andreferenceIEEECSLogNumberTVCG-0028-0306.DigitalObjectIdentifierno.10.1109/TVCG.2007.1008.

1077-2626/07/$25.00ß2007IEEE

M

“getridofthisobscuringhighlight,makesomeshadowstorevealroughtexturehere,butfillintheshadowsthere”),butquicklydeterminingwhatcombinationsoflightstouse,wheretoplacethem,andhowtoadjustthemforthebestoveralleffectcantakeyearstomaster.

Oursystemaddressesthesedifficultiesinprovidingstudio-qualitylighting.Weareinterestedincomputer-assistedlightingfordesktop-sizedstationaryobjects.Wewantlightingthataccuratelyrevealstheshape,texture,andthemostvisuallymeaningfulfeaturesofthephotographedobject.Weshowthatsketch-guidedoptimization,asimpleoff-the-shelflightsource,andideasfromimage-basedlightingcansubstantiallyreducetheequipment,patience,andskillrequiredforhigh-qualityphotography.Theuserplacestheobjectonatableand(eithermanuallyorautomatically)spendsafewminutestotakeasimpleseriesofphotographs(<200),eachlitfromdifferent,butover-lapping,directions.Aftergatheringthesephotos,thesystempresentsaninitialimageresult.Theuserthensimplysketcheshis/herlightingpreferencesdirectlyonthisimage,andthesystemcomputesanewphotothatbestmatchesthesketchintheleastsquaressense.Thesketch-guidedlightingadjustmentsrapidlyconvergetoattrac-tivelylitphotographicresults.

Weenvisiontwoapplicationsinparticular:tohelpmuseumcuratorsquicklygatherrelightabledigitalphoto-graphicarchivesandtohelpcasualphotographerscreateoneortwowell-litphotos(forexample,forpostingoneBay)withaslittlelightingequipmentaspossible.Thispaperoffersthreecontributions:1)Weextendexistingimage-basedlightingideastoreducetherequiredequip-menttoasinglelightsourceandsinglecamera.2)Wereplacetrial-and-errorrepositioningofmultiplelightsourceswithoptimizationandon-screenpainting.3)Weeliminate,orgreatlyreduce,theneedforexhaustivehigh-dynamic-range(HDR)photographytoreducecapturetime.Theresultisanovelandinexpensivesystemthatanovice

PublishedbytheIEEEComputerSociety

MOHANETAL.:TABLETOPCOMPUTEDLIGHTINGFORPRACTICALDIGITALPHOTOGRAPHY653

Fig.1.Acomparativeoverviewofoursystem.(a)Lightinginaphotographer’sstudioiscomplicatedandslowtosetupbutproducesgoodresults.(b)Usingasteerablespotlightandanuncalibratedenclosure,(c)oursystemproduceslightingofcomparablequality.Inastudio(a),anexperiencedphotographermanuallypositionsmultiplelightstoproduceprofessionalresults.Thisprocessreliesonthephotographer’sknowledgeandexperiencetoproduceahigh-qualityresult.Oursystem(b)usesafoam-coreenclosureandsteerablelighttophotographtheobjectautomaticallylitfrommanydifferentdirections.Thesephotographsformthebasisofanoptimizationthatproducescomparablequalitylighting(c)whenauserinteractivelysketchesthedesiredlighting.

canusetointuitivelydescribeandobtainthedesiredlightingforaphotograph.

2RELATEDWORK

Pioneeringworkinimage-basedlighting[8],[13],[9],[18]offeredpromisingapproachesthatcouldhelpwiththephotographiclightingproblem.Unfortunately,mostsys-temsrequiredtoomanyprecisemeasurementsandadjust-mentsforday-to-dayuseoutsidethelaboratory.Suchprecisionwasnecessaryforambitiousgoalssuchasrecoveringshape,bidirectionalreflectancedistributionfunction(BRDF),andappearanceunderarbitraryviewingandlightingconditions.Forthemuchsmallerbutmorewidespreadproblemofphotographiclighting,weneedamethodandapparatusthatrequireslesstime,expense,andcomplexity,yetallowsuserswhoarenotlightingexpertstoquicklyfindthelightingtheywant.

Findingexpressivelightingforascenehaslongbeenrecognizedasadifficultproblemincomputergraphicsandseveralpapershavealreadyexploredoptimizationforlightplacementandotherparameters[26],[15],[22],[7],[27].Someofthesesystemsalsousedpaintinginterfacestospecifythedesiredlightinginscreenspace[26],[22],[23],andweuseasimilarapproachtomakelightingforphotographymoreintuitive.ThesystembyShackedandLischinski[27]wasevenabletoprovidefullyautomaticlightingbyapplyingimagequalitymetrics.However,allofthesesystemsrequire3Dinformationunavailableinourphotographicapplication.

AsindicatedinFigs.1band3,oursystemusesasimpleenclosureforautomaticdatagathering.Severalcompaniesalsoofferphotographicenclosures,butonlytoachieveverysoftlighting;theydonothelpuserssolvelightplacementproblemsasoursdoes.Thesesystemsincludediffusivetents[21],photoboxes[19],andtranslucentbacklitplatformswithanarrayofindividuallydimmedlightsources[4].Image-basedmethodshavealsobeenusedtoperformarbitraryrelightingofwell-measuredobjects.Mostmeth-ods,includingours,performrelightingusingaweightedsumofdifferentlylitbasisimagesasdescribedbyNimeroffetal.[20],whoobservedthatlightandmaterialsinteract

linearly.However,prioreffortsusedmuchmoreelaborateequipmentbecausetheirgoalsdifferedfromours.Theseincludedmeasurementofa4DsliceofthereflectancefieldofthehumanfacebyDebevecetal.[8],museumartifactsmeasuredbyarotating-armlightstagebyHawkinsetal.[13],aningeniousbutelaboratesystembyDebevecetal.[9]forreal-timevideoplaybackandmeasurementoflightfields,adomeofelectronicflashunitstogatherrelightabletexturesbyMalzbenderetal.[16],afree-formlightstagetoenableportablegatheringoflight-fielddatawithsomecalibrationbyMasselusetal.[17],andfull4DincidentlightmeasurementsbyMasselusetal.[18].Inallofthesecases,datagatheringrequiredeithercustomizedequipmentorcollectiontimesmuchlongerthanwouldbepracticalforroutinephotographiclightingtasks.ArecentworkbyFuchsetal.[12]gathereddataquicklywithasinglehandheldlight.Unlikeeitherourenclosedspotlightorourhandheldlightwithattachedreflector,theyaimedtheirlightattheroom’swhite-paintedwallsandestimatedincidentilluminationanglesusingalightprobe(chromesphere)ineachphoto.NovelBayesiantechniquesfurtherincreasedthedirectionalselectivityoftheirrelightingbuttradedselectivityforincreasednoise.

Threerecentsystemsalsoofferednovelsketch-guidedrelightingfrombasisimages.Akersetal.[2]usedaroboticlight-positioninggantrytogatherpreciselylitimagesand,likeus,alsoprovidedapaintinginterfacetoguiderelightingchoices,but,unlikeus,theyusedspatiallyvaryingweightsthatcouldproducephysicallyimpossiblelighting.Agarwalaetal.[1]usedsketch-guidedgraph-cutsegmentationcoupledwithgradientdomainfusiontoseamlesslymergeseveralphotographsformanydifferentpurposes.Theydemonstratedmergingdifferentlylitphotographstocreatenovelilluminationconditions.Thoughtheirinteractionschemeworkedwellforasmallnumberofimages($10),itmaybeimpracticalforthehundredsofimagesrequiredforcompletecontroloverlightingdirections.Also,theirsystemdoesnothingtohelptheuserwithlightplacementandmayproducephysically

´[3]usedaDebevec-unrealizableresults.AnrysandDutre

stylelightstagewithabout40fixed,low-poweredlightsourcesandapaintinginterfacetoguidelightingchoices.

654IEEETRANSACTIONSONVISUALIZATIONANDCOMPUTERGRAPHICS,VOL.13,NO.4,JULY/AUGUST2007

Fig.2.Allpossiblelightinganglesparameterizedbylightpositionanddirection.Pointlightsources(ontheleft)resultinmultiplehardshadows,whereasoverlappingarealightsources(ontheright)canbeusedtosimulatealargerlightsource.

Theiroptimizationchoseonlythelightintensities,andusersstillchosethelightplacements.Also,theirpoint-likelightsourcescouldcausemultipleshadowsandhighlightsintheoptimizedresult,whichmaybeundesirableforarchivalpurposes.TheirdatacapturetimeisalsohighsincetheycaptureHDRphotosforeverylightlocation.

Oursystemdiffersbyusingoptimizationtodeterminethelightpositions,thenumberofcontributingbasisimages,andtheirweights.Byusingoverlappedareaemittersinsteadofpoint-likelights,wecanconstructsoftshadowapproximationswithoutapparentmultipleshadows.Werequirearound5minutesformanualorautomaticimagecaptureandafewmoreminutesofuserinteractiontogetthefinalresult.Theequipmentrequiredisminimalandportable,andeverythingneededforthemanualversioncanfitinabackpack.

3

SIMPLIFICATIONSAND

CONSTRAINTS:

2DLIGHTING

Ourgoalistocreateasystemthathelpsusersfindgoodlightingforascenebutallowsthemtointeractivelyspecifytheresultratherthanthelightsneededtoachieveit.Toconstructwell-litscenes,studiophotographersmakecon-sistentchoicesaboutwhichtypesoflightstouse,wheretoplacethem,andhowtoadjustthem.Ourstreamlined,image-basedmethodmakessimilarchoices(seeFig.1).Wedonotseektobuildacompletecalibrated4Ddatasettoreconstructallformsofilluminationbutonlytogatherenoughdatatocomputeawell-litphotograph.

TheobservationofNimeroffetal.[20]thatlightsandmaterialsinteractlinearlymeansthat,ifafixedcameramakesanimageIifromafixedscenelitonlybyalightLi,thenthesamescenelitbymanyiLiwillmakeanimageIout¼Plightsscaledbyweightswimage,asiftheiwiIi.Adjustingweightsletsus“relight”theweightsmodulatethelightsratherthantheimages.AswecollectmoreimagesIi,wecansimulatemorelightingpossibilities.Severalcom-monpracticesinstudiolightingcanguideusinchoosingtherightnumberofimagestogather.

First,professionalphotographerscommonlychooselampswithbroad,nearlyuniformbeamsoflight,oftenwithareflectorandlenstohelpdirectmorelightforward.Second,theyadjustlightplacementanglescarefully,butnottheirdistances.Thedistancetothelightaffectsforeshorten-ingofshadowshapes,buttheseeffectsaresubtleandrarely

Fig.3.Themoving-headspotlightsetup.Theobjectanddiscolightarebothenclosedinawhitefoambox,withthecameralookinginthroughawindowintheenclosurewallfarthestfromthelight.

noticedinstillimages.Third,theyadjustlightsourceareastocontrolshadowsoftnessversussharpness.Lightsources(or,moreaccurately,theshadowstheyform)become“softer”byincreasingtheangularextentasmeasuredfromthelitobject.Fourth,theyseekoutlightingarrangementsthatproduceasimplesetofshadowsandhighlightsthatbestrevealtheobject’sshape,position,andsurfacequalities.Theyavoidcomplexoverlappedshadows,appar-entlackofshadowsduetooverlysoftlight,andcontrastextremesfromlargespecularhighlightsorverydarkshadows.Simplershadowsusuallymeanfewerlightsand,thus,fewerbasisimages.

Accordingly,weuseonecommerciallyavailablemov-ing-headlightplacedinsideanenclosurewhosewallsreflectlightbackontheobject.Thisgivesushundredsofcontrollable“soft”arealightsourcesatamoderatedistance(typicallyaboutonemeter)aroundtheobject.Thisisincontrastwiththemuchsharperpoint-likesourcesfoundinmanyearlierapproaches.Anarealightsourceformssoftershadowsand,usingweightswitoblendmultipleover-lappedarealights,itformslargersourcesthatavoidthedirectionalaliasingcausedbyblendinglightfromagroupofseparatepoint-likesources.Thisdirectionalaliasingappearsasvisuallydisturbingmultiplesharpshadows,asshowninFig.2.

Notethatwedonotneedtoknoweithertheexactlightpositionsortheirabsoluteintensitiesforourtechnique;weselectweightswiandimagesIibytheirabilitytomatchthelightingtargetimagesausersketchesforus.However,reliableaddressabilityandrepeatabilityishelpfulbecauseitallowsustoselectivelyreturntoanypreviouslightsourcepositionandgatheranymultiple-exposureHDRimagesneededforthefinaloutputimage.Theseareusefulif,forexample,allbasisimageshaveashinyspecularhighlightthatisoverexposed.Fortunately,nearlyallcommerciallyavailabledigitallyaimedspotlightsofferveryaccurate,finelyaddressable,andrepeatablesteering.

WeavoidtheuseofHDRphotographswherepossible,asthesetypicallyrequiremultiplecalibratedexposuresandcomputationtomergethem.Instead,werelyonthecamera’sautomaticexposureadjustmentstocaptureimagessuitableforinteractivelightingdesign.Inaddition,diffusedinterreflectionswithinourenclosurearoundthephoto-graphedobjectserveasanambientlightsourcethatreduces

MOHANETAL.:TABLETOPCOMPUTEDLIGHTINGFORPRACTICALDIGITALPHOTOGRAPHY655

Fig.4.Anillustrativesampleofimagesgatheredfromonecapturesessionusingthemoving-headspotlightandfoam-coreenclosure.

scenecontrastbyaddinggentlenearlyuniformilluminationinalldirections,lighteningshadowsthatmightotherwiseappearasfeaturelessblackinthescene.Thus,asmen-tioned,weresorttoHDRcapturemethodsonlyforbasisimageswithlargeoverexposedregions.Underexposedregionscanbesafelyignored,astheircontributionstotheresultingimagearealreadyinvisibleandfurtherreducedbyweightswi 1:0.

Formally,arbitraryexternalilluminationis4Dforadesktopscene,butwecapturea2Dsubsetdefinedbyilluminationangles.Followingcommonstudiopractice,weignoredistancefromtheobjecttothelightsource.Thespotlight’sprojectedspotontheinteriorofourenclosureapproximatesasmallextendedarealightsourceilluminat-ingthephotographedobject.Ratherthanrecreatearbitrary4Dincidentlightfields,weuseweightedsumsofphotographs(“basisimages”),eachtakenwithdifferent2Dilluminationangles.

Weconstructahigh-qualityuser-guidedpictureinthreesteps.First,thesystemorusercapturesasetofphotoswithafixedtripod-mountedcamera.Eachofthese“basisimage”photoscapturesthesceneilluminatedfromslightlydifferentangles,denselysamplingtheaddressableanglesonoursinglelightsource.Second,theuseriterativelyobservesadisplayedimageandpaintsitinteractivelywithsimplelightenanddarkenoperationstoindicatedesiredlightingchanges,formingthe“target”image.Third,thesystemperformsoptimization;itfindsweightswiforeachbasisimagesuchthattheirweightedsumformsthebestmatchtothetargetimageintheleastsquaressense.Thisweightedsumofbasisimagesbecomesthenewdisplayedimage.Ifdissatisfiedwiththeresult,usersmaycontinuetosketchandrefinethisresulttocreateanewtargetimageforanotheriterationofsketch-guidedimprovement.

4STEP1:IMAGEGATHERING

4.1Automatic:SpotlightandEnclosure

Freedfromphotometricandangularcalibrationrequire-mentsasdiscussedinSection3,asimpleandcost-effectivecontrolledlightsourceismucheasiertobuild.Weplacetheobjectandagimbal-mountedmoving-headspotlightinsideanenclosureofalmostanyconvenientsize,shape,andmaterial.Thepowerfulcomputer-controlledspotlightpi-votstoanydesiredpanandtiltanglewithgoodrepeat-abilityð Æ0:5󰀅Þtolightanydesiredspotinsideourenclosure.Theenclosureactsasareflectorandeffectivelyprovidesacontrollable2Darealightsourcethatsurroundstheobjectfromanydirectionabovethegroundplane.Thesizeandshapeoftheenclosureisalmostirrelevantaslongasthespotlightiscloseenoughtotheobjecttokeepthe

parallaxlowandthelightispowerfulenoughforthecameratogetareasonableexposure.

Webuilta1Â1Â1:5m3-sizedboxofwhite1=200foam-coreboardasourenclosureandchoseaninexpensivemoving-headspotlighttoplaceinsideit.The150WAmericanDJAutoSpot150discolightcantilt270degrees,canpan540degrees,andincludesninecolorfilters,gobos,andseveralotherfunfeatures.ComputercontrolbytheDMX512protocoliseasytoprogram,andweusedthecommerciallyavailableSoundLightUSBDMXcontroller.Ourfoam-coreenclosureresemblesahemicubesuspendedjustaboveapairoftables.Weplacethegimballightonaseparateshorttablethatlowersitsrotationcentertotheplaneofanadjacenttallertableholdingthephotographedobject,asshowninFig.3.Separatetablesreducevibration,permitgimbalanglestoapproximatehemisphereangles,andsafelyseparatetheobjectfromtheswivelinglamp.Weplacethecamerabehindasmallopeningcutintheenclosurewallontheendfarthestfromthelightsource.Thesystemgathersimagesrapidlyandautomatically.ThroughtheDMX512controller,wedirectthegimballighttoscantheupperhemisphereoflightaimingdirectionsinequal-angleincrementsaswerecordindividualcomputer-triggeredphotographsusingthecamera’sautoexposuresettings.Wegatheranexhaustivesetofabout130imageswithoverlappinglightingdirections.OurCanonPowershotG3camerarequiredapproximately3secondsperphoto(includingautoexposureestimationandtheactualexpo-sure,butnotthedatatransfertime).AsampleofimagesfromonesuchcapturesessionisshowninFig.4.

Tothebestofourknowledge,nootherimage-basedlightingworkexploitsthesemovableorcontrollablelights.Enclosedpivotinglightsretainmanyadvantagesofthemoresophisticatedlightingsystems,avoidmultiplesharpshadows,offervariable“softness”byspotsizeadjustment,andaremuchsimplerandcheapertoconstruct.Ofcourse,theydonoteasilyprovideaccuratelightingdirectioncalibrationorpoint-lightillumination,butthesefeaturesarenotneededforourgoals.

Afterrecording,welinearizeeachcapturedframebyapplyingthecamera’sinverseresponsecurve,recoveredbythemethodofDebevecandMalik[10],andconstructalow-resolution(320Â240)luminance-onlyimagesetforuseintheoptimizationstep(Section6).Linearresponseensuresthatweightedsumsofwholeimagesareaccuraterepre-sentationsofphysicallyrealizablelighting.

4.2Manual:HandheldLightandReflector

Evenafoam-coreboxandamoving-headspotlightareimpracticaltocarryaroundeverywhere.However,the

656IEEETRANSACTIONSONVISUALIZATIONANDCOMPUTERGRAPHICS,VOL.13,NO.4,JULY/AUGUST2007

Fig.5.Handhelddatagatheringusingalightwithanattachedfoam-corediffusereflector.

“free-formlightstage”[17]showedthatitispossibletogathercalibratedimagesetssuitablefor2Drelightingwithnothingmorethanfoursmalllight-probe-likespheres,adigitalcameraandtripod,ahandheldpoint-lightsource,possiblybatterypowered,andapproximately30minutesoftimetotakeseveralhundreddigitalphotographs.Althoughitmeetstheambitiousgoalofincidentlightfieldcapture,themethodmighttaxanyone’spatiencetorecordmorethanjustafewitems.Wepresentafasterandsimplervariantthatservesourpurposebetter.

Intheautomaticmethodjustdescribed,repeatablelightsourcepositioningallowsustoacquireHDRimagesselectively(asdiscussedinSection6.1),afteruserlightingchoicestelluswhichimageswillmakesignificantcontributionstothedisplayedoutput.However,ifweeitherignoreoverexposedspecularhighlightsortakethetimetorecordHDRimagesduringcapture,thenrepeatabilityisnotneeded;wecanuseahandheldlightsourceforoursystemaswell.AsshowninFig.5,weuseasmall250Whandheldlightintendedfortelevisionnewscameras,attachedtoadiffusereflectormadeoffoamcore,andlimitthebeamwidthwithbarndoorstoformawell-definedarealightsource.

Togatherallphotos,weholdthelightoutstretchedandmovethelightonahemispherecenteredabouttheobjectinwhatwecallour“lightgatheringdance.”Wesamplethehemisphereoflightingdirectionsbyapolar-coordinatescan(󰀂istheazimuthand󰀃istheelevation,asshowninFig.2)in󰀂firstorderasthecameratakessequentialphotographs.ANikonD70cameratakesasteadystreamofphotosataboutthreeframespersecond,usingautoexposureforeachframe.Theuserstandsfacingtheobjectandholdsthelightatarms’lengthwhilemovingthelampinanarcthatpassesdirectlyovertheobject.Theusermovesthelampfromonesideofthetabletotheother,scanningby󰀄radiansin󰀃axiswithconstant󰀂,andthenaturalalignmentoftheirshouldershelpsaimthelight’scenterlinedirectlyattheobject.Aftereachpassovertheobjectwiththelight,theuserstepssidewaystochangethe󰀂angleforthenextscanandmakesenoughofthesepassestocover0 󰀂<󰀄radians.Inpractice,theusercanbemorecarelesswiththelight,aslongasthehemisphereoflightiswellsampledandtheimagesarenotoverexposed.Aftertheimagecapture

Fig.6.Ascreenshotoftheuserinterfaceforcreatingtargetlightingconditions.Thetargetimagethattheusersketchesonisthetopimage,shownincolor.Thepanelshowstheresultofanypreviousiterationsaswell.Thebottomblack-and-whiteimageisthemask;itistypicallynotdisplayedunlesstheusersodesires.Themaskisoptionalandselectsthosepixelsforwhichthelightingoptimizationwillbeperformed,thuspotentiallyimprovingperformance.Thecontrolboxesontheleftgovernwhattypeofadjustmenttothetargetimageandmaskarebeingmade,forexample,contrastordodgeandburn.Thecontrolpanelalsogovernsthebrushsizeandothercontrolsforoptimization,outputassembly,andsaving.

danceiscomplete,wedownsampleallimagesandproceedwiththesketch-guidedlightingdesignasbefore.

Wefindthisprocesstobequitesimpleandpleasingand,inlessthan4minutes,wecangatherahigh-qualitybasisphotosetof120-150images.Anexperiencedusermightnotneedtoscanthewholehemispherebutcanquicklyilluminatejusttheregionswheretheyknowtheyneedcomputedlightsources.

5STEP2:TARGETIMAGESKETCHING

Afterthelightgatheringstep,userscaniterativelyspecifyandrefinephotographiclightingbysketchingonatargetimage.Thisgray-scaleimage(seeexamplesinFig.8)approximatesthefinaloutputimagetheuserwouldliketosee.TheuserinterfaceforthisapplicationisshowninFig.6.

Ineditingthetargetimage,theuserstartsoffeitherwithasimpleluminanceimageofthetarget,agraywash(forexample,uniformgrayorlightgrayfadingtodarkgrayacrosstheimage),orthepreviousiteration’sresult.Theuserthencarriesoutaseriesoflightenanddarkenoperationsinthedifferentregionsoftheimagetoapproximatethedesiredresults.Theusercanpaintareasofluminanceontheimagedirectly,useadodge-and-burncontrol,oradjustthecontrast.Theprocessissimpleandintuitiveandtakesafewminutesatmost.Anoptionalmask,whichisabilevelimagethatindicateswhichpixelsintheimagewillbesubjecttotheoptimization,maybeapplied.Themaskcansignificantlyspeeduptheoptimizationprocessiflighting

MOHANETAL.:TABLETOPCOMPUTEDLIGHTINGFORPRACTICALDIGITALPHOTOGRAPHY657

changesinsomeareasoftheimageareunimportant.Themaskisgeneratedusingasimplemodelthataddsapixelpaintedoverbytheusertothemaskandcanbemanuallychangedifdesired.

6STEP3:OPTIMIZATION

AND

DISPLAY

Givenatargetimage,theoptimizationfindsweightswiforeachdownsampledimagethatprovidesthebestmatchtothetargetimage.Wetakeaconstrainedleastsquaresapproach.LetNbethenumberofimagesinthebasisset,eachofsizemÂn.Wefind

minkAwÀtk2

w

ð1Þ

suchthat

0 wi 1

foralli,wherewistheN-dimensionalvectorofweights,AisanðmÂnÞÂNmatrixofbasisimages(eachbasisimageistreatedasavector),andtistheðmÂnÞvectorrepresentingthetargetimagepaintedbytheuser.Weightsareconstrainedsincenegativeweightswouldgivenegativelighting,whichisnotphysicallyrealizable,andweightsgreaterthanone,althoughphysicallyrealizable,introduceexcessivenoiseintotheimages.Ingeneral,thisoptimizationschemethereforeassumesthattheimagesettakenbythecameraisnot,asaset,underexposed.

Theresultisaleastsquaresoptimalmatchtothesuppliedtargetimage.Astheobjectivefunctionisquad-ratic,weightsforimageswithweakcontributionsarerapidlydriventozero.Inourexperience,thenumberofsignificantnonzeroweightsisconsistentlysmall(forexample,5-15).Thus,thenumberofimagesneededforthefinallightingsolutionisalsotypicallysmall.

Weemployabound-constrainedlimited-memoryvari-ablemetric(BLMVM)methodforconstrainedoptimization[5]thatispartoftheTAOoptimizationpackage[6].Onatargetimagesizeof170Â227,with150basisimages,theoptimizationtakesapproximately70secondsona1GHzPentium4.Iftheuserchangesthetargetimageandrepeatstheoptimization,thistimeusingonlytheimagescorre-spondingtothenonzeroweights,thentheoptimizationtakeslessthan1secondandisinteractive.Thisinteractivitycomeswiththeassumptionthattheuserhaschosenalloftheproperfeaturestoilluminateinthefirstsketchandwillnotstartaddingsmallspecularitiestotheimage.Totestthereasonablenessandutilityofthisfeature,weconductedpilotstudieson12naiveusersofthesystem.Inthesestudies,mostusersfoundtheinitialdelayacceptableandfeltthattheycouldadequatelyrefinetheimageastheywantedusingthereducedbasissetatinteractiverates.Theoptimizationroutineweemployisnotguaranteedtofindaglobalminimum.However,wetestedouroptimiza-tionfunctionagainstasimplex-basedsimulatedannealingalgorithmwithrandomrestarts[24].Forseveralruns,withaslowannealingschedulesothattheoptimizationtookover30hoursandwithrandominitialconditions,thesimulatedannealingalgorithmconvergedclosetothesameresidualcostandwiththesamesetofweights.WealsotookthesolutionproducedbytheBLMVMmethodandran

simulatedannealingwiththatasaninitialcondition.TheBLMVMmethodwasneverimproveduponinanytest.Therefore,ourtestsindicatethatdifferentoptimizationstrategiesconvergetothesamevalue,givingusconfidenceintherobustnessofoursolution.

Afterfindingthewiweights,weapplythemtothelinearizedbasisimages,thenreapplythecameraresponsefunctiontodisplayapreviewoftheoutputimage.Theuserthenhastheoptionofreplacingthetargetwithagray-scaleversionofthisresultandcanrepeatthesketchingandoptimizationcycleuntilsatisfiedwiththecolorpreviewoftheoutputimage.

6.1AssemblingtheDisplayedImage

Usersaregenerallysatisfiedwiththeresultsafteriteratingthroughthesketchingprocessafewtimes.However,insomecases,somebasisimagesmaycontainoverexposedareaswithbrightspecularhighlightsthatmightcauseartifactsintheresult.TheseresultscansometimesbeimprovedbyusingHDRphotographsfortheselightingdirections.

Iftheautomaticmethodwasusedtocapturetheimageset,thenHDRphotographsforthefinalimagesetcanbecapturedbecausethelightingpositionisknownandrepeatable.Wealsoassumethattheobjecthasnotmovedorcanbeplacedinarepeatableconfiguration.Asbefore,weconstructalinearoutputimageasaweightedsumofbasisimages,usingtheweightsdeterminedbytheoptimizationtomatchthetargetimage.Finally,wereapplythecamera’sresponsefunctiontothelinearoutputimagetogetthedesiredresult.ThisHDRtechniqueisnotpracticalforthemanualhandheldacquisitioncasesincelightpositionsarenoteasilyrepeatable.

7ISSUES

OF

CONTROLLABILITY

Thissectionpresentsananalysisofhowmuchcontroltheuserhasoverthelighting,giventhattheeventuallightingwillbearealizablecombinationoflightswiththephysicalconstraintsofthemechanismsdescribedinSection4.Weconductedpilotstudieswithnaiveusersinwhichtheywereabletoachievetheresultstheywanted,but,togainabetterunderstandingofthisissue,weusedthePrincipalCompo-nentAnalysis(PCA)[14]tocomputetheeigenimagesofthelightingdataset[28].Priorworkanalyzingvariabilityinlightingconditionshasshownthat,forconvexLambertianobjects,fiveeigenimagessufficetocapturealmostallofthelightingvariation[25],andthat,whendiffusespecularitiesandslightlymorecomplexgeometryareadded,onlyoneortwoadditionaleigenimagesareneeded[11].Ourworkfallsintoasimilarcategoryasthispriorworkinthatweassumethattheviewpointisheldconstant.Differentlyfromthepriorwork,though,isourinterestinsmallareasofhighspecularity,small-scalecastshadows,andotherfeaturesgeneratedbyirregularlyshapedpossiblyshinyobjects.Thesefeaturesareonesthatusersmaywishtovaryandcontroltoimprovetheaestheticfeaturesofthefinalresult.Thesesmallvariationswilltypicallynotbecapturedinthefirstfiveorseveneigenimages.

ExampleeigenimagesthatallowustointerprettheseresultsareshowninFig.7.Tounderstandhowtointerpret

658IEEETRANSACTIONSONVISUALIZATIONANDCOMPUTERGRAPHICS,VOL.13,NO.4,JULY/AUGUST2007

Fig.7.Threeeigenimagesforseverallightingdatasetspresentedascolorimages.Thetoprowshowsthefirsteigenimageinred,thesecondinblue,andthethirdingreen.Thebottomrowshowsselectedothereigenimagesthatindicatethefinenessofthecontrolthatcanbeachievedwithoursystem.Notethatsimilarlycoloredregionsindicateregionsoflikecovariation,whichimpliesthatalightingchangerequestedbytheuserinthisregionwillaffectthelightingintherestoftheregion.Thefinalimageshowsthenormalizedprincipalvaluesshowntogetherwiththebroken-stickvaluesforthecastledataset.Theprincipalvalueisshowninblueandthebrokenstickvalueasadashedline.

theseimages,notethatabrightpixelinaneigenimageindicatesapixelofhighcovariationwhenconsideredwithotherpixelsinthateigenimage.Sinceeacheigenimageisagray-scaleimage,wecanvisualizethreeeigenimagesbycreatingacolorimagewiththefirsteigenimageastheredchannel,thesecondasthegreenchannel,andthethirdasthebluechannel.ExamplesofthisconstructionforthefirstthreeeigenimagesofvariousdatasetsareshowninthetoprowofFig.7.Theresultingsimilarlycoloredregionsshowusareaswhere,ifwechangethelightinginthatregion,therestofthelightingwillcovaryaswell.Thus,thetoprowshowsthevariationinlightingandcontrolthatwidecoarsepaintingofthetargetislikelytoachieve.So,forexample,intheimageofthedragonshowninthefirstrowofFig.7,auserlighteningordarkeningthebackgroundintheupperleftcorneroftheimagewilllikelybelighteningordarkeningthesnoutofdragonsincebothregionsarered.Thefirstthreeeigenimagesshowonlythecoarsestlightingvariationthatcanbeachieved.Oursystemalsoperformsreasonablywellforsmallspecularvariationsandlightinginself-occludedregions,ascanbeseenintheeigenimagesinthebottomrowofFig.7.Theseimagesshowselectedothereigenimagesfromseveralofourdatasets.Inthem,wecansee,forexample,thatausercancontrolsmallspecularhighlightsontheearsofthedragon(showninblue)andsomespecularareasonthewings(showninred).Also,thethreeturretsinthecastleobjectcanbegivenanicevariationoflightingacrossthem,asshowninthecastlepictureonthesecondrow.Thus,theseeigenimagesindicatehowthelightingislikelytochangewithbroadbrushstrokesinthetargetimagesandalsothefinenesswithwhichlightingcanbecontrolled.

Thecoloredeigenimagesgiveaninterestingvisualiza-tionofthethreemajorregionsofcovariationbutdonotgiveacompletecharacterizationofthecontrollability,sincetherearemorethanthreeprincipalcomponents.Consistentwiththeworkcitedabove,wefindthatthefirstthreeprinciplecomponentsaccountforover80percentofthelightingvariabilityasevaluatedusingthebrokenstickmodel[14]tocomputethenumberofsignificantdirections.Therearebetweenfourandsevensignificantlightingcovariationsforthedatasetspresentedinthispaper.Theprincipalvalues

correspondingtotheprincipalcomponentsforthecastledatasetisshowninthebottomrowofFig.7.Thisfigureshowsthenormalizedprincipalvalues(theeigenvaluesdividedbytheirtrace)withthebrokenstickvalue.Aninterestingquestionwehavenotfullyexplorediswhetheroursystemcouldperformthelightingoptimizationsatisfactorilyinalowerdimensionalspace.Ensuringthatthelightingremainsphysicallyrealizableafteroptimizationinthisspaceseemstobethemainchallenge.Also,webelievethatthedatapresentedinthispaperrepresentsawidesampleofinterestingobjects,andamorecompletecharacterizationwouldrequirethegeometryofanitemanditsmaterialpropertiestobeconsidered.

8RESULTS

EveryimageinFig.8showsresultsfromoursketch-guidedlightingsystem.Boththemoving-headlightandthehandheldmethodsareequallysuccessfulatcreatingarbitrarycleanlylitimagesofdesktop-sizedobjects.Thedatasetsgatheredbyeithermethodaresufficientlydensetoallowflexiblelightingdesign.Additionally,oursystemyieldsreasonableresultsevenwhenpresentedwithunrealistictargetsorhighlyreflectiveobjects.

Fig.8ademonstratesauserinteractionsequencewiththesystem.Startingfromauniformgray-scaleimageasthetarget,theuserguidestheoptimization,iterativelyimprov-ingthetargetuntilshegetsthedesiredoutput.Fig.8bshowshowsimpleapproximatesketchingonthetargetimagecangiveaninterestingside-lightingeffect.Fig.8cshowshowthehighlightcanbringouttheunderlyingtextureinasurface.

Fig.8dshowslightingforahighlyspecularobject.Goodlightingforsuchsmooth,highlyreflectiveobjectsisalwaysdifficult,asthelightsourceitselfisvisibleinthereflection.Oursystemproducesresultssimilartothetargetimagewithoutlargeobjectionablesaturatedregions.Infuturesystems,wemayhidetheenclosureseamsbyconstructingwidesmoothroundedcornersresemblingaphotographer’s“cyc.”

MOHANETAL.:TABLETOPCOMPUTEDLIGHTINGFORPRACTICALDIGITALPHOTOGRAPHY659

Fig.8.Sampleresultsfromourlightingdesignsystem.Eachimageisshownwithitstargetimage.(a)Sequenceshowingsuccessivesketching/optimizationiterationstogetthedesiredlighting.Thefirstresultusesaconstantgray-scaletarget,whereastheothersusepreviousresultsasstartingpointsforthetargetimage.(b)Strategicplacementofhighlightsinthetargetresultsinaninterestingside-litimage.(c)Positioningofhighlightsrevealstheunderlyingtextureinthesurface.(d)Lightingahighlyspecularobjectbyforcingthebackgroundtobedark.(e)Targetresultsinimagesuggestingilluminationfromtheright.(f)Datacapturedbythehandheldmethod.Theimageontheleftusesasmoothgray-scalegradientasthetargetimage.

660IEEETRANSACTIONSONVISUALIZATIONANDCOMPUTERGRAPHICS,VOL.13,NO.4,JULY/AUGUST2007

Fig.9.Comparingdiffuseversusdirectlightingfortwodifferentmodelsshows(a)thetargetimageandmaskforeachset,respectively,(b)thediffuselylitset,(c)thedirectlylitset,and(d)theimageproducedfromthecombinedset.

Fig.8fshowsresultsfromthemanualhandheldmethod.Thedatagatheringtimewasunder3minutesandtheresultsarecomparabletothemoving-headlightmethod.Althoughthehandheldmethodmaynotbepracticalforphotographingalargecollectionofobjects,itcanbeaninvaluabletoolforwell-litphotographyinthefieldorasasimpleconsumer-levelimplementation.

Fig.9showsacomparisonofadiffuselightsourceversusamorefocusedoneonbothadiffuseandaspeculardataset.Thediffuseobjectisasmallmodelofacastle,whichcontainsinterestingself-shadows.Thespecularobjectisacrystalandgolddragonsculpture.Thediffuselylitsetswerecapturedforbothobjectsusingthehandheldlightingdeviceandthemorefocusedonesbyremovingthediffusereflectorandshiningthelightdirectlyonthemodels.TargetimagesandmasksareshownincolumnFig.9a.Foreachtargetimage,weusedthreebasissets:thediffuse(Fig.9b),thedirect(Fig.9c),andbothbasissetstogether(Fig.9d).Notsurprisingly,thedirectresultbasissettendstoresultinmoreshadowartifactsbutsharperhighlightsthanthediffusebasissets.Thecombineddatasetblendsthefacetsofboth.

9DISCUSSION

AND

FUTUREWORK

Thispapertakestheproblemofgoodlightingfordesktopphotographyandfindsasimpleandpracticalsolutionusingimage-basedrelightingtechniques.Ourcomputeduser-guidedresultsarealwaysphysicallyrealizable;wecouldcapturethesameimageinasinglephotographinastudiosettingbypositioningseveralcustomizedlightsourcesaroundtheobject.Oursystem’soutputisalwaysaweightedsumofseveralradiometricallylinearizedphotos,andtheseweightsarealwaysbetweenzeroandone.Weightslessthanzeromightrequireimpossiblenegativelightsources,andweightsgreaterthanonewouldamplifynoisecontributionsfromtheoriginalphotographs,especiallyindeeplyshadowedregions.

Weightsbetweenzeroandonepresenttheirowndifficulties,evenafterradiometriclinearization.Ifthedynamicrangeofthephotographedsceneexceededthecamera’smeasurementabilities,thentheminimumandmaximumpixelvalues(forexample,0or255)arenolongermeaningful.Zero-valuedpixelsscaledbyoneorlessremainvalidbecausetheseradiometriccontributionsarealreadytoosmalltoseeinthefinalimage,butscaledcontributions

MOHANETAL.:TABLETOPCOMPUTEDLIGHTINGFORPRACTICALDIGITALPHOTOGRAPHY661

from“clipped”specularhighlights(forexample,255)willunderestimatethescenecontributionstotheoutputimage.Oursystemavoidstheseproblemsintwoways.First,interreflectionswithinourdiffuseenclosureprovidedimalmost-uniformambientlightingthatactsasa“fill”light,illuminatingotherwiseshadowedregionstoreducethescene’soveralldynamicrange.Second,oncetheoptimiza-tionstephaschosenthebasisimagesfortheoutputimage,wechecktheseimagesforclippedvaluesthathavenoradiometricmeaning.Inthesecases,weneedatleastonemorephotographatalowerexposuresettingtomeasurethehighlightvalues,buttheaddressabilityandrepeat-abilityofoursteerablespotlightletsusrecreatethenecessarylightingandtaketheadditionalphotographsneededtocompletetheoutputimage.

Theabilitytohavelargearealightsourcesiscrucialforphotographinghighlyspecularobjects.Lightsourcesizealsoaffectsthesharpnessofshadowsandhighlights.Oursystemhastheuniqueadvantagethatlargerarealightsourcescanbesimulatedbycombiningpictureswithoverlappinglightsources.Wecouldextendouroptimiza-tiontopenalizeeachdistinctlightsourcecluster,thuspreventingdisjointhighlights.Thesoftnessofthelightcanalsobecontrolledbyvaryingthebeamwidthbetweenapointsourceandalargeareasourceasitquicklysweepsoverthehemisphereoflightingdirections.Moreadvancedmoving-headspotlightsusuallyprovidecontrollablespotsizessuitableforthispurpose.

Eventhoughthesystemisaimedprimarilyatnonpro-fessionalphotographers,afewsimpleadditionscanmakeitaflexibletoolforacreativeexperttoexperimentwithdifferentlightingdesignsmoreeasily.Toolstodirectlytweakthelightpositionandsizeonavirtualhemispherearoundtheobjectmightalsoaidexpertusers.Moresophisticatedimage-basedmeasurementsmightalsobeachievablewhilemaintainingthesimplicityandeleganceofthesystem.Forexample,wecouldcalibrateouradhocenclosuretomeasureincidentlightanglesasafunctionofgimbalangleseasilyfromasetofaimingimagesofachromesphereor“lightprobe.”Combinedwithsurfacenormals,suchcalibrationmightsufficeforimage-basedestimatesofBRDF.

ACKNOWLEDGMENTS

TheauthorsthankJingjingMeng,XingHu,andPinRenfortheirhelpintheearlystagesoftheproject;KelliJohnsonandNathanMatsudafortheirhelpwiththeconstructionoftheenclosure;VincentMasselusandAmyGoochfortheirhelpwithFig.2;andElizabethSewardforherhelpinthepilotstudies.ThisresearchwasfundedinpartbyUSNationalScienceFoundationGrants0238062,0237621,and0535236.Anyopinions,findings,andconclusionsorrecommenda-tionsexpressedinthismaterialarethoseoftheauthorsanddonotnecessarilyreflecttheviewsofthesponsors.

REFERENCES

[1]A.Agarwala,M.Dontcheva,M.Agrawala,S.Drucker,A.Colburn,B.Curless,D.Salesin,andM.Cohen,“InteractiveDigitalPhotomontage,”Proc.SIGGRAPH,vol.23,no.3,pp.294-302,2004.[2]

D.Akers,F.Losasso,J.Klingner,M.Agrawala,J.Rick,andP.Hanrahan,“ConveyingShapeandFeatureswithImage-BasedRelighting,”Proc.IEEEVisualization,2003.

[3]F.AnrysandP.Dutre

´,“ImageBasedLightingDesign,”Proc.FourthIASTEDInt’lConf.Visualization,Imaging,andImageProcessing,2004.

[4]AstronSystems,“OrbiculightLightingSystem,”http://www.astronsys.com,2007.

[5]

S.J.BensonandJ.More

´,“ALimitedMemoryVariableMetricMethodinSubspacesandBoundConstrainedOptimizationProblems,”TechnicalReportANL/MCS-P909-0901,Math.andComputerScienceDivision,ArgonneNat’l[6]

S.J.Benson,L.C.McInnes,J.More

´Laboratory,2001.

,andJ.Sarich,“TAOUserManual(Revision1.7),”TechnicalReportANL/MCS-TM-242,Math.andComputerScienceDivision,ArgonneNat’lLaboratory,2004,http://www.mcs.anl.gov/tao.

[7]A.C.Costa,A.A.Sousa,andF.N.Ferreira,“LightingDesign:AGoalBasedApproachUsingOptimisation,”Proc.EurographicsWorkshopRendering,vol.10,pp.317-328,1999.

[8]P.Debevec,T.Hawkins,C.Tchou,H.-P.Duiker,W.Sarokin,andM.Sagar,“AcquiringtheReflectanceFieldofaHumanFace,”Proc.SIGGRAPH,pp.145-156,2000.

[9]P.Debevec,A.Wenger,C.Tchou,A.Gardner,J.Waese,andT.Hawkins,“ALightingReproductionApproachtoLive-ActionCompositing,”Proc.SIGGRAPH,pp.547-556,2002.

[10]P.E.DebevecandJ.Malik,“RecoveringHighDynamicRangeRadianceMapsfromPhotographs,”Proc.SIGGRAPH,vol.31,pp.369-378,1997.

[11]

R.Epstein,P.W.Hallinan,andA.L.Yuille,“5Æ2EigenimagesSuffice:AnEmpiricalInvestigationofLow-DimensionalLightingModels,”Proc.IEEEWorkshopPhysics-BasedModelinginComputerVision,pp.108-116,1995.

[12]M.Fuchs,V.Blanz,andH.-P.Seidel,“BayesianRelighting,”RenderingTechniquesandP.Dutre

´2005:EurographicsSymp.Rendering,K.Bala

eds.,pp.157-1,July2005.[13]T.Hawkins,J.Cohen,andP.Debevec,“APhotometricApproachtoDigitizingCulturalArtifacts,”Proc.Conf.VirtualReality,Archeology,andCulturalHeritage,pp.333-342,2001.[14]I.Jolliffe,PrincipalComponentAnalysis.Springer,1986.

[15]J.K.Kawai,J.S.Painter,andM.F.Cohen,“Radioptimization:GoalBasedRendering,”Proc.SIGGRAPH,pp.147-154,1993.

[16]T.Malzbender,D.Gelb,andH.Wolters,“PolynomialTextureMaps,”Proc.SIGGRAPH,pp.519-528,[17]V.Masselus,P.Dutre

´2001.

,andF.Anrys,“TheFree-FormLightStage,”Proc.13thEurographicsWorkshop[18]V.Masselus,P.Peers,P.Dutre

´Rendering,pp.247-256,2002.

,andY.D.Willems,“Relightingwith4DIncidentLightFields,”Proc.SIGGRAPH,vol.22,no.3,pp.613-620,2003.

[19]MKDigital,“PhotoLightBox,”http://www.mkdigitaldirect.com,2007.

[20]J.S.Nimeroff,E.Simoncelli,andJ.Dorsey,“EfficientRe-RenderingofNaturallyIlluminatedEnvironments,”Proc.FifthEurographicsWorkshopRendering,pp.359-373,1994.

[21]Photek,“DigitalLighthouse,”http://www.photekusa.com,2007.[22]P.PoulinandA.Fournier,“PaintingSurfaceCharacteristics,”Proc.EurographicsWorkshopRendering,pp.160-169,1995.

[23]P.Poulin,K.Ratib,andM.Jacques,“SketchingShadowsandHighlightstoPositionLights,”Proc.Conf.ComputerGraphicsInt’l,pp.56-63,1997.

[24]W.H.Press,S.A.Teukolsky,W.T.Vetterling,andB.P.Flannery,NumericalRecipesinC,seconded.CambridgeUniv.Press,1992.[25]

R.Ramamoorthi,“AnalyticPCAConstructionforTheoreticalAnalysisofLightingVariabilityinImagesofaLambertianObject,”IEEETrans.PatternAnalysisandMachineIntelligence,vol.24,no.10,pp.1322-1333,Oct.2002.

[26]C.Schoeneman,J.Dorsey,B.Smits,J.Arvo,andD.Greenburg,“PaintingwithLight,”Proc.SIGGRAPH,pp.143-146,1993.

[27]R.ShackedandD.Lischinski,“AutomaticLightingDesignUsingaPerceptualQualityMetric,”ComputerGraphicsForum,vol.20,no.3,pp.215-226,2001.

[28]

M.TurkandA.Pentland,“FaceRecognitionUsingEigenfaces,”J.CognitiveNeuroscience,vol.3,no.1,1991.

662IEEETRANSACTIONSONVISUALIZATIONANDCOMPUTERGRAPHICS,VOL.13,NO.4,JULY/AUGUST2007

AnkitMohanreceivedtheBEdegreefromtheNetajiSubhasInstituteofTechnology,India,in2001.HeisagraduatestudentintheDepart-mentofElectricalEngineeringandComputerScienceatNorthwesternUniversity.Hiscurrentandlatentinterestsincludecomputationalphoto-graphy,image-basedlighting,globalillumina-tion,operatingsystems,andphotography.

ReynoldBaileyreceivedthebachelor’sdegreeincomputerscienceandmathematicsin2001fromMidwesternStateUniversityandthemaster’sdegreeincomputersciencein2004fromWashingtonUniversity,St.Louis,whereheiscurrentlyaPhDstudent.Hisresearchinter-estsincludeappliedperceptioningraphicsandvisualizationandnonphotorealisticrendering.

JonathanWaitereceivedtheBSdegreeincomputersciencefromVanderbiltUniversityin2006.HeiscurrentlywithPathfinderTherapeu-tics,Nashville,Tennessee.

JackTumblinreceivedthemaster’sdegreeinelectricalengineeringin1990andthePhDdegreeincomputersciencefromtheGeorgiaInstituteofTechnologyin1999.HehasbeenanassistantprofessorintheDepartmentofElectricalEngineeringandComputerScienceatNorthwesternUniversitysince2001,aftertwoyearsofpostdoctoralstudyatCornellUniversityandayear-longvisitatMicrosoftResearchfrom1994-1995.Hecofoundedthe

flightsimulatorcompanyIVEXCorp.Heholdsninepatents.HeisamemberoftheIEEE.

CindyGrimmreceivedtheBAdegreeincomputersciencefromtheUniversityofCalifor-niaBerkeleyin1990andtheMAandPhDdegreesincomputersciencefromBrownUni-versityin1992and1995,respectively.SheisaprofessorintheDepartmentofComputerScienceandEngineering,WashingtonUniver-sity,St.Louis,wheresheworksinthefieldofcomputergraphics.SheisamemberoftheIEEE.

BobbyBodenheimerreceivedthePhDdegreeinelectricalengineeringfromtheCaliforniaInstituteofTechnology,Pasadena.Heiscur-rentlyanassistantprofessorofcomputerscienceatVanderbiltUniversity,Nashville,Tennessee.Previously,hewasavisitingre-searcheratMicrosoftResearch,Redmond,Washington,andapostdoctoralfellowattheGeorgiaInstituteofTechnology,Atlanta.Hisresearchfocusesoncomputergraphics,com-puteranimation(especiallyhumanfigureanimation),andvirtualenvironments.HereceivedtheUSNationalScienceFoundation(NSF)FacultyEarlyCAREERDevelopmentAward.HeisaseniormemberoftheIEEE.

.Formoreinformationonthisoranyothercomputingtopic,pleasevisitourDigitalLibraryatwww.computer.org/publications/dlib.