Comparing Simplification and Image-Based Technique

ComparingSimplificationandImage-BasedTechniquesfor3DClient-ServerRenderingSystemsW.Pasman,F.W.JansenDelftUniversityofTechnologyABSTRACTAmathematicalmodelispresentedforcomparinggeometricandimage-basedsimplificationmethods.Geometricsimplificationreducesthenumberofpolygonsinthevirtualobjectandimage-basedsimplificationreplacestheobjectwithanimage.Ourmodelintegratesandextrapolatesexistingaccuracyestimates,enablingthecomparisonofdifferentsimplificationmethodsinordertochoosethemostefficientmethodinagivensituation.Themodelcomparesdatatransferandrenderingloadofthemethods.Bytesizeandexpectedlifetimeofsimplificationsarecalculatedasafunctionofthedesiredvisualqualityandthepositionandmovementoftheviewer.Anexampleresultisthatintypicalviewingandrenderingconditionsandforobjectswitharadiusintheorderofonemeter,impostertechniquescanbeusedatviewingdistancesabovefifteenmeters.Belowthat,simplifiedpolygonobjectsarerequired,andbelowonemeterdistancethefull-resolutionvirtualobjecthastoberendered.Anelectronicversionofthemodelisavailableontheweb.INDEXTERMS:realtimerendering,dynamicgeometrysimplification,imposters,resourceload,thinclient,mathematicalmodel1INTRODUCTIONVariousgeometricandimage-basedsimplificationtechniques,suchassimpleimposters[Schaufler96,Maciel95],meshedimposters[Sillion97,Decoret99]andsimplifiedpolygonmodels[Schroeder92]havebeendevelopedtoadaptthecomplexityofascenetotheavailablebandwidthandcapacityoftherenderingengineandnetwork.Thesesimplificationandimpostertechniquesoftenpreserve,foragivenviewpoint,crucialaspectsoftheobjectssuchascontourandfrontimage,whilesacrificinggeometricaccuracyforotherviewpoints.Thisintroducesthenotionoflifetimeforthesimplifiedrepresentation:astheviewpointchanges,thevisualdistortionwillgrowandwillrunthesimplificationobsolete.Therepresentationthenwillhavetoberefreshedorreplacedwithanotherformofgeometricsimplification.Weassumethattherenderingandsimplificationareseparatedandcanbecharacterisedasaserver-clientarchitecture,wheretheserverholdsthecompletescenemodeldatabaseandsuppliestherenderingclientwithascenedescriptionofreducedsizewiththeappropriatelevelofdetail.Thissimplifiedscenecanberenderedinrealtimewhilemakingaslittlesacrificestotheimagequalityaspossible.Asimilarclient-serversetupcanbefoundinweb-basedsystemswherethedatatransfercapacityandlatencyofthecommunicationlink,aswellasthetemporalvalidityofpartsofthemodelimposesadditionalconstraintsonthemodelcomplexity[Campbell98,Kim96].Ourfocusisonthinclients,suchasmobilephonesandmobileaugmentedrealitysystems[UbiCom01],whereitisessentialtomanagetheCPU-memory-andcommunicationloadontheclient.Nevertheless,ourmodelalsoappliestofat-clientsetups.Wewillrestricttheobjectrepresentationstosingle-resolutionmesheswithmulti-resolutiontextures,inordertoavoidexcessivecomplexityinthemodelling.Intuitively,simpleimpostersarecheapesttotransmitandrender,andremaincorrectrelativelylongiftheobjectisfaraway.Butfornearbyobjectstheviewpointchangesrelativelyrapidlyandquicklyoutdatestheimposter.Inthatcaseasimplifiedpolygonobjectispreferable.Forobjectsatmoderatedistance,themeshedimposterseemstobethemostappropriate.SeeFigure1.Shadeetal.[Shade98]suggestsasimilaruseofvarioussimplificationmethods.ObserverOutsideviewingcone:lowerqualityNearby:fullpolygonobjectsFurtheraway:meshedimposter,imagewithdepthFaraway:simpleimposterViewingconeFigure1.Varioussimplificationmethodsaremosteffectiveatdifferentdistances.Ourpurposeis(1)tocomparethedifferentsimplificationmethodsand(2)toselecttheappropriaterepresentationsdynamicallygivenlimitedresources.Forthis,wepresentamodelforthesesimplificationmethodsthatrelatestheloadonthecommunicationlinkandtheloadontherenderingenginetothevisualaccuracy.Westartwithadiscussionofpreviouswork.Insection3wegiveanoverviewofourapproach,andlisttheassumptionswemade.Insection4wederivedistortionformulasforthevarioussimplificationmethods.Theseformulasareinvertedinsection5,toestimatetherequirednumberofpolygonsandthesizeofthetexturesgivenadesiredaccuracy.Thesesizescanthenbeconvertedintoabytesize.Thecommunicationloadisthenestimatedbydividingtheestimatedbytesizebytheplannedlifetimeofthevirtualobjects.Insection6,resultsareplottedforthevarioussimplificationmethods,tocomparetheloadonthecommunicationlinkandontherenderingengine.Apartialvalidationofthemodelisgiveninsection7.Conclusionsaresummarisedinsection8.2PREVIOUSWORK2.1Latencyinthin-clientrenderingarchitecturesThissectionreviewslatencyissueswithrespecttothin-clientarchitectures.Latencyisthetimelagbetweenthemomenttheuserchangeshisviewpointandthemomenttheimagesforhisnewviewpointareactuallypresentedtohim.Latencyisaproblemwithmanyvirtualreality(VR)andaugmentedreality(AR)applications,asitwillcauseincorrectplacementandfloatingofvirtualobjects,andmayevencausesimulatorsickness.ForVR,latenciesof60to150msareverynoticeable[Aliaga97]andforcardriving30msseemsthemaximum[Padmos92].Predictionhasbeenusedtoalleviatetheseproblems[Azuma95].Howeversuchpredictionsgiveacceptableresultsonlywhenthepredictedtimeisverysmall,andproblemsbecomeworseasheadsetsbecom