By scattered data we mean an arbitrary set of points in which carry scalar quantities ie a scalar 64257eld in dimensional parameter space In contrast to the global nature of the leastsquares 64257t the weighted local ap proximation is computed eithe ID: 3054
Download Pdf The PPT/PDF document "An AsShortAsPossible Introduction to the..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Figure1:Fittingbivariate,quadraticpolynomialsto2Dscalarelds:thetoprowshowsthetwosetsofninedatapoints(seetext),thebottomrowshowstheleastsquarestfunction.Thecoefcientvectors[c1;:::;c6]Tare[0:834;0:25;0:75;0:25;0:375;0:75]T(leftcolumn)and[0:334;0:167;0:0;0:5;0:5;0:0]T.MethodofNormalEquations.Foradifferentbutalsoverycom-monnotation,notethatthesolutionforcinEqn.(3)solvesthefollowing(generallyover-constrained)LSE(Bc=f)intheleast-squaressense264bT(x1)...bT(xN)375c=264f1...fN375;(5)usingthemethodofnormalequationsBTBc=BTfc=(BTB)1BTf:(6)PleaseverifythatEqns.(4)and(6)areidentical.2WLSApproximationProblemFormulation.Intheweightedleastsquaresformulation,weusetheerrorfunctionalJWLS=åiq(k xxik)kf(xi)fik2foraxedpoint x2Rd,whichweminimizeminf2Õdmåiq(k xxik)kf(xi)fik2;(7)similarto(1),onlythatnowtheerrorisweightedbyq(d)wherediaretheEuclidiandistancesbetween xandthepositionsofdatapointsxi.Theunknowncoefcientswewishtoobtainfromthesolutionto(7)areweightedbydistanceto xandthereforeafunctionof x.Thus,thelocal,weightedleastsquaresapproximationin xiswrittenasf x(x)=b(x)Tc( x)=b(x)c( x);(8)andonlydenedlocallywithinadistanceRaround x,i.e.kx xkR.WeightingFunction.Manychoicesfortheweightingfunctionqhavebeenproposedintheliterature,suchasaGaussianq(d)=ed2 h2;(9)wherehisaspacingparameterwhichcanbeusedtosmoothoutsmallfeaturesinthedata,see[Levin2003;Alexaetal.2003].AnotherpopularweightingfunctionwithcompactsupportistheWendlandfunction[Wendland1995]q(d)=(1d=h)4(4d=h+1):(10)Thisfunctioniswelldenedontheintervald2[0;h]andfurther-more,q(0)=1,q(h)=0,q0(h)=0andq00(h)=0(C2continuity).Severalauthorssuggestusingweightingfunctionsoftheformq(d)=1 d2+e2:(11)Notethatsettingtheparameteretozeroresultsinasingularityatd=0,whichforcestheMLStfunctiontointerpolatethedata,aswewillseelater.Solution.AnalogoustoSection1,wetakepartialderivativesoftheerrorfunctionalJWLSwithrespecttotheunknowncoefcientsc( x)åiq(di)2b(xi)[b(xi)Tc( x)fi]=2åi[q(di)b(xi)b(xi)Tc( x)q(di)b(xi)fi]=0;wheredi=k xxik.Wedividebytheconstantandrearrangetoobtainåiq(di)b(xi)b(xi)Tc( x)=åiq(di)b(xi)fi;(12)andsolveforthecoefcientsc( x)=[åiq(di)b(xi)b(xi)T]1åiq(di)b(xi)fi:(13)Obviously,theonlydifferencebetweenEqns.(4)and(13)aretheweightingterms.Noteagainthough,thatwhereasthecoef-cientscinEqn.(4)areglobal,thecoefcientsc( x)arelocalandneedtoberecomputedforevery x.IfthesquarematrixAWLS=åiq(di)b(xi)b(xi)T(oftentermedtheMomentMatrix)isnonsingular(i.e.det(AWLS)6=0),substitutingEqn.(13)intoEqn.(8)providesthetfunctionf x(x).GlobalApproximationusingaPartitionofUnity(PU).Byt-tingpolynomialsatj2[1:::n]discrete,xedpoints xjinthepa-rameterdomainW,wecanassembleaglobalapproximationtoourdatabyensuringthateverypointinWiscoveredbyatleastoneapproximatingpolynomial,i.e.thesupportoftheweightfunctionsqjcenteredatthepoints xjcoversWW=[jsupp(qj):Properweightingoftheseapproximationscanbeachievedbycon-structingaPartitionofUnity(PU)fromtheqj[Shepard1968]jj(x)=qj(x) ånk=1qk(x);(14)whereåjjj(x)1everywhereinW.Theglobalapproximationthenbecomesf(x)=åjjj(x)b(x)Tc( xj):(15) ANumericalIssue.Toavoidnumericalinstabilitiesduetopossi-blylargenumbersinAWLSitcanbebenecialtoperformthettingprocedureinalocalcoordinatesystemrelativeto x,i.e.toshift xintotheorigin.Wethereforerewritethelocaltfunctionin xasf x(x)=b(x x)Tc( x)=b(x x)c( x);(16)theassociatedcoefcientsasc( x)=[åiq(di)b(xi x)b(xi x)T]1åiq(di)b(xi x)fi;(17)andtheglobalapproximationasf(x)=åjjj(x)b(x xj)Tc( xj):(18)3MLSApproximationandInterpolationMethod.TheMLSmethodwasproposedbyLancasterandSalka-uskas[LancasterandSalkauskas1981]forsmoothingandinterpo-latingdata.Theideaistostartwithaweightedleastsquaresfor-mulationforanarbitraryxedpointinRd,seeSection2,andthenmovethispointovertheentireparameterdomain,whereaweightedleastsquarestiscomputedandevaluatedforeachpointindividu-ally.Itcanbeshownthattheglobalfunctionf(x),obtainedfromasetoflocalfunctionsf(x)=fx(x);minfx2Õdmåiq(kxxik)kfx(xi)fik2(19)iscontinuouslydifferentiableifandonlyiftheweightingfunctioniscontinuouslydifferentiable,seeLevinswork[Levin1998;Levin2003].SoinsteadofconstructingtheglobalapproximationusingEqn.(15),weuseEqns.(8)and(13)(or(16)and(17))andcon-structandevaluatealocalpolynomialtcontinuouslyovertheen-tiredomainW,resultingintheMLStfunction.Aspreviouslyhintedat,using(11)astheweightingfunctionwithaverysmalleassignsweightsclosetoinnityneartheinputdatapoints,forcingtheMLStfunctiontointerpolatetheprescribedfunctionvaluesinthesepoints.Therefore,byvaryingewecandirectlyinuencetheapproximatimg/interpolatingnatureoftheMLStfunction.4ApplicationsLeastSquares,WeightedLeastSquaresandMovingLeastSquares,havebecomewidespreadandverypowerfultoolsinComputerGraphics.Theyhavebeensuccessfullyappliedtosurfacerecon-structionfrompoints[Alexaetal.2003]andotherpointsetsurfacedenitions[AmentaandKil2004],interpolatingandapproximatingimplicitsurfaces[Shenetal.2004],simulating[Belytschkoetal.1996]andanimating[M¨ulleretal.2004]elastoplasticmaterials,PartitionofUnityimplicits[Ohtakeetal.2003],andmanyotherresearchareas.In[Alexaetal.2003]apoint-set,possiblyacquiredfroma3Dscanningdeviceandthereforenoisy,isreplacedbyarepresenta-tionpointsetderivedfromtheMLSsurfacedenedbytheinputpoint-set.Thisisachievedbydown-sampling(i.e.iterativelyre-movingpointswhichhavelittlecontributiontotheshapeofthesurface)orup-sampling(i.e.addingpointsandprojectingthemtotheMLSsurfacewherepoint-densityislow).Theprojectionproce-durehasrecentlybeenaugmentedandfurtheranalyzedintheworkofAmentaandKil[AmentaandKil2004].Shenet.al[Shenetal.2004]useanMLSformulationtoderiveimplicitfunctionsfrompolygonsoup.Insteadofsolelyusingvalueconstraintsatpoints(asshowninthisreport)theyalsoaddvalueconstraintsintegratedoverpolygonsandnormalconstraints.ReferencesALEXA,M.,BEHR,J.,COHEN-OR,D.,FLEISHMAN,S.,LEVIN,D.,ANDT.SILVA,C.2003.Computingandrenderingpointsetsurfaces.IEEETransactionsonVisualizationandComputerGraphics9,1,315.AMENTA,N.,ANDKIL,Y.2004.Deningpoint-setsurfaces.InProceed-ginsofACMSIGGRAPH2004.BELYTSCHKO,T.,KRONGAUZ,Y.,ORGAN,D.,FLEMING,M.,ANDKRYSL,P.1996.Meshlessmethods:Anoverviewandrecentdevel-opments.ComputerMethodsinAppliedMechanicsandEngineering139,3,347.FRIES,T.-P.,ANDMATTHIES,H.G.2003.Classicationandoverviewofmeshfreemethods.Tech.rep.,TUBrunswick,GermanyNr.2003-03.LANCASTER,P.,ANDSALKAUSKAS,K.1981.Surfacesgeneratedbymovingleastsquaresmethods.MathematicsofComputation87,141158.LEVIN,D.1998.Theapproximationpowerofmovingleast-squares.Math.Comp.67,224,15171531.LEVIN,D.,2003.Mesh-independentsurfaceinterpolation,toappearin'ge-ometricmodelingforscienticvisualization'editedbybrunnett,hamannandmueller,springer-verlag.M¨ULLER,M.,KEISER,R.,NEALEN,A.,PAULY,M.,GROSS,M.,ANDALEXA,M.2004.Pointbasedanimationofelastic,plasticandmelt-ingobjects.InProceedingsof2004ACMSIGGRAPHSymposiumonComputerAnimation.OHTAKE,Y.,BELYAEV,A.,ALEXA,M.,TURK,G.,ANDSEIDEL,H.-P.2003.Multi-levelpartitionofunityimplicits.ACMTrans.Graph.22,3,463470.PRESS,W.,TEUKOLSKY,S.,VETTERLING,W.,ANDFLANNERY,B.1992.NumericalRecipesinC-TheArtofScienticComputing,2nded.CambridgeUniversityPress.SHEN,C.,O'BRIEN,J.F.,ANDSHEWCHUK,J.R.2004.Interpolatingandapproximatingimplicitsurfacesfrompolygonsoup.InProceedingsofACMSIGGRAPH2004,ACMPress.SHEPARD,D.1968.Atwo-dimensionalfunctionforirregularlyspaceddata.InProc.ACMNat.Conf.,517524.WENDLAND,H.1995.Piecewisepolynomial,positivedeniteandcom-pactlysupportedradialbasisfunctionsofminimaldegree.AdvancesinComputationalMathematics4,389396. Figure2:TheMLSsurfaceofapoint-setwithvaryingdensity(thedensityisreducedalongtheverticalaxisfromtoptobottom).ThesurfaceisobtainedbyapplyingtheprojectionoperationdescribedbyAlexaet.al.[2003].ImagecourtesyofMarcAlexa.