Irani P Anandan and D Weinshall Dept of Applied Math and CS The Weizmann Inst of Science Rehovot Israel iraniwisdomweizmannacil Microsoft Research One Microsoft Way Redmond WA 98052 USA anandanmicrosoftcom Institute of Computer Science Hebrew Univ ID: 29831
Download Pdf The PPT/PDF document "From Reference Frames to Reference Plane..." 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.
FromReferenceFramestoReferencePlanes:Multi-ViewParallaxGeometryandApplicationsM.Irani,P.Anandan,andD.WeinshallDept.ofAppliedMathandCS,TheWeizmannInst.ofScience,Rehovot,Israel,MicrosoftResearch,OneMicrosoftWay,Redmond,WA98052,USA,InstituteofComputerScienceHebrewUniversity91904Jerusalem,IsraelThispaperpresentsanewframeworkforanalyzingthege-ometryofmultiple3Dscenepointsfromuncalibratedimages,basedondecomposingtheprojectionofthesepointsontheimagesintotwostages:(i)theprojectionofthescenepointsontoa(realorvir-tual)physicalreferenceplanarsurfaceinthescene;thiscreatesavirtualimageonthereferenceplane,and(ii)there-projectionofthevirtualimageontotheactualimageplaneofthecamera.Thepositionsofthevirtualimagepointsaredirectlyrelatedtothe3Dlocationsofthescenepointsandthecameracentersrelativetothereferenceplanealone.Alldependencyontheinternalcameracalibrationparametersandtheori-entationofthecameraarefoldedintohomographiesrelatingeachimageplanetothereferenceplane.Bi-linearandtri-linearconstraintsinvolvingmultiplepointsandviewsaregivenaconcretephysicalinterpretationintermsofgeometricrela-tionsonthephysicalreferenceplane.Inparticular,thepossibleduali-tiesintherelationsbetweenscenepointsandcameracentersareshowntohavesimpleandsymmetricmathematicalforms.Incontrasttotheplane+parallax(p+p)representation,whichalsousesareferenceplane,theapproachdescribedhereremovesthedependencyonareferenceim-ageplaneandextendstheanalysistomultipleviews.Thisleadstosim-plergeometricrelationsandcompletesymmetryinmulti-pointmulti-viewduality.Thesimpleandintuitiveexpressionsderivedinthereference-planebasedformulationleadtousefulapplicationsin3Dsceneanalysis.Inparticular,simplertri-focalconstraintsarederivedthatleadtosimplemethodsforNewViewSynthesis.Moreover,theseparationandcompactpackingofcameracalibrationandorientationintothe2Dprojectiontransformation(ahomography)allowsalsopartialreconstructionusingpartialcalibrationinformation.Keywords:Multi-pointmulti-viewgeometry,uncalibratedimages,newviewsynthesis,dualityofcamerasandscenepoints,plane+parallax,trilinearity. M.IraniissupportedinpartbyDARPAthroughARLContractDAAL01-97-K-0101 1IntroductionTheanalysisof3Dscenesfrommultipleperspectiveimageshasbeenatopicofconsiderableinterestinthevisionliterature.Giventwocalibratedcameras,theirrelativeorientationscanbedeterminedbyapplyingtheepipolarconstrainttotheobservedimagepoints,andthe3Dstructureofthescenecanberecoveredrelativetothecoordinateframeofareferencecamera(referredtohereasthereferenceframe e.g.,see[13,6]).ThisisdonebyusingtheepipolarconstraintandrecoveringtheEssentialMatrixEwhichdependsontherotationbetweenthetwocameras.ConstraintsdirectlyinvolvingtheimagepositionsofapointinthreecalibratedviewsofapointhavealsobeenderivededIfthecalibrationofthecamerasisunavailable,thenitisknownthatre-constructionisstillpossiblefromtwoviews,butonlyuptoa3Dprojectivetransformation[4].Inthiscasetheepipolarconstraintstillholds,buttheEssen-tialMatrixisreplacedbytheFundamentalMatrix,whichalsoincorporatestheunknowncameracalibrationinformation.The3Dscenepoints,thecameracentersandtheirimagepositionsarerepresentedin3Dand2Dprojectivespaces(usinghomogeneousprojectivecoordinates).Inthiscase,thereferenceframereconstructionmayeitherbeareferencecameracoordinateframe[8],orasde-nedbyasetof5basispointsinthe3Dworld[14].Acompletesetofconstraintsrelatingtheimagepositionsofmultiplepointsinmultipleviewshavebeende-rived[5,15].Alternatively,givenaprojectivecoordinatesystemspeciedby5basispoints,thesetofconstraintsdirectlyrelatingtheprojectivecoordinatesofthecameracenterstotheimagemeasurements(in2Dprojectivecoordinates)andtheirdualconstraintsrelatingtotheprojectivecoordinatesofthe3Dscenepointshavealsobeenderived[2,20].Alternatively,multipleuncalibratedimagescanbehandledusingtheplane+parallax(P+P)approach,whichanalyzestheparallaxdisplacementsofapointbetweentwoviewsrelativetoa(realorvirtual)physicalplanarsurfaceinthescene[16,12,11].Themagnitudeoftheparallaxdisplacementiscalledtherelative-anestructurein[16].[12]showsthatthisquantitydependsbothontheHeightanditsdepthrelativetothereferencecamera.Sincetherelative-ane-structuremeasureisrelativetoboththereferenceframe)andthereferenceplane(through),werefertotheP+Pframe-workalsoasthereference-frame+reference-planeformulation.TheP+Phasthepracticaladvantagethatitavoidstheinherentambiguitiesassociatedwithestimatingtherelativeorientation(rotation+translation)betweenthecameras;thisisbecauseitrequiresonlyestimatingthehomographyinducedbythereferenceplanebetweenthetwoviews,whichfoldstogethertherotationandtranslation.Also,whenthesceneisat,thematrixestimationisunstable,whereastheplanarhomographycanbereliablyrecovered[18].Inthispaper,weremovethedependencyonthereferenceframeoftheanal-ysisofmulti-pointmulti-viewgeometry.Webreakdowntheprojectionfrom3Dto2Dinto2operations:theprojectionofthe3Dworldontothe2Dreference,followedbya2Dprojectivetransformation(homography)whichmaps thereferenceplanetotheimageplane.Giventhevirtualimagesformedbytheprojectionontothereferenceplane,wederivealgebraicandgeometricrela-tionsinvolvingtheimagelocationsofmultiplepointsinmultipleviewsinthesevirtualimages.Thepositionsofvirtualimagepointsaredirectlyrelatedtothe3Dlocationsofthescenepointsandthecameracentersrelativetothereferenceplanealone.Alldependencyontheinternalcameracalibrationparametersandtheorientationofthecameraarefoldedintohomographiesrelatingeachimageplanetothereferenceplane.WeobtainastructuremeasurethatdependsontheheightsofthescenepointsrelativetothereferenceplaneInthispaper,wederiveacompletesetdualrelationshipsinvolving2and3pointsin2and3views.Onthereferenceplanethemulti-pointmulti-viewgeometryissimpleandintuitive.Theserelationsaredirectlyrelatedtophys-icalpointsonthereferenceplanesuchastheepipoleandthedual-epipoleoleWeidentifythesepoints,andalsotwonewentitiescalledthetri-focallinedualtrifocal-linewhichareanalogoustotheepipoleandthedual-epipolewhenconsideringthree-viewandthree-pointgeometriesonthereferenceplane.Structuressuchasthefundamentalmatrixandthetrilineartensorhavearathersimpleformanddependonlyontheepipoles,andnothingelse.Thesymmetrybetweenpointsandcamerasiscomplete,andtheycanbesimplyswitchedaroundtogetfromtheepipolargeometrytothedual-epipolargeometry.Thesimpleandintuitiveexpressionsderivedinthereference-planebasedformulationinthispaperleadtousefulapplicationsin3Dsceneanalysis.Inparticular,simplertri-focalconstraintsarederived,andtheseleadtosimplemethodsforNewViewSynthesis.Also,theseparationandcompactpackingofcameracalibrationandorientationintothe2Dprojectiontransfor-mation(ahomography)thatrelatestheimageplanetothereferenceplane,leadstopotentiallypowerfulreconstructionandcalibrationalgorithms.Forinstance,basedonminimalpartialdomaininformation,partialcalibrationandpartialreconstructioncanbeachieved.Thisisalsobrieydiscussedinthispaper.Theremainderofthispaperisorganizedasfollows:Section2introducesournotations,anddescribesthetwo-viewgeometricandalgebraicconstraints(bi-linearityandparallax)inthereferenceplanerepresentation.Section3describesduality(betweenscenepointsandcameracenters)onthereferenceplane.Sec-tion4examinestherelationsinvolving3viewsandthecorrespondingdualre-lations.Section5discussesapplicationsofthisrepresentationandshowsinitialresultsforoneparticularapplication,namelynew-viewsynthesis.2TwoViewGeometryontheReferencePlaneFigure1illustratesthetwostagedecompositionoftheimageformationprocess.Figure1ashowstheprojectionofonescenepointfromtwocameracentersontothereferenceplane.Figure1bshowsthere-projectionfromtheplanetooneofthecameraimageplanes(thereferenceframe).Inthisandinallsubsequentguresinthispaper,weadoptthefollowingnotations:denotesscenepointsin3D,denotescameracenters;i,j,kareindicesusedforscenepoints(e.g., (a) ispitpts PiCsCtHiHsHtPlane pis H CsCt Reference Reference Image Plane pitpts Fig.1.Thereferenceplanerepresentation:(a)theprojectionofthepointsontothereferenceplaneitself,removingthedependencyontherefer-enceimageplane.istheepipole,andtheredlineistheepipolarline.(b)re-projectionofthereferenceplaneimageontoareferenceimageframe(camerat).)andr,s,tareindicesusedforcameracenters(e.g.,).Also,denotestheprojectionofthescenepointthroughcameracenter.Itistheintersectionoftheraywiththereferenceplane.Similarlyistheintersectionofwiththereferenceplane.Wedeneasthevirtual-imagesofonthereferenceplanefromcamerasrespectively.Wedenetheintersectionofastheepipole.Weusetodenotetheepipole.NotethatthelocationoftheepipoleonthereferenceplaneisisindependentoftheorientationsandtheinternalcalibrationparametersoftheToderivethealgebraicconstraintsinvolvingmultiplepointsinmultipleviews,wedeneacoordinatesystem(x,y,Z)relativetothereferenceplane,where(x,y)areparalleltoisperpendiculartoit.Forpointsonthereferenceplane,wedene=1,forotherpointswedene+1,denotestheheight(i.e.,theperpendiculardistance)ofthepointfromtheplane.Thus,),where+1,denotesthe3Dcoordi-natesofthescenepoint.Similarly,where+1,and,where+1.Thepointsonthereferenceplanearetheintersectionsofthelines,andwiththereferenceplane HiHtHiytHtyi HiHt1pis=xityit1=HixsHsxi HiHsHiysHsyi HiHs1(1)pts=xtsyts1=HsxtHtxs HsHtHsytHtys Notethattheexpressionsgivenabovedonotinvolveanyofthecamerainternalcalibrationparametersortheorientationsoftheimageplanes.Alsonotethat thereisonlyasingleepipole,whichisunlikethecaseofthereference-framebasedformulation,whichinvolvestwoepipoles,oneoneachimageframe.Thepointsonthereferenceplanearerelatedtotheircor-respondingpointsonanimageplane(e.g.,areferenceimage)viaasingle2Dprojectivetransformation,whichisthehomographybetweenthatimageplaneandtheplane.Figure1bshowsthere-projectionontothereferenceimagethepointsp,p,andaretheprojectionsoftheimagepoints,andtheepipolerespectively.Therearetwobasicresultsconcerningtwoviewsofapointasobservedonthereferenceplane.Therstistheexpressionfortheparallaxonthereferenceplane,andthesecondisthebilinearconstraintinvolvingthetwoimagelocationsofthescenepointandtheepipole.Thesearedescribedbelow.ParallaxontheReferencePlane:GiventheexpressionsinEquations1and2,itcanbeeasilyveriedthat Notethatthisexpressionforparallax(Equation3)involvesonlytheheightsofthescenepointandofthecameracentersrelativetothereferenceplane.Itdoesnotincludeanyquantitiesrelativetoanyofthecameracoordinatesystems(e.g.,thereferenceframe)suchasasbefore.Also,theparallaxdoesnotdependonthex,ylocationsofeitherthecameracentersorthescenepointTheBilinearConstraint:Equation3impliesthat,andarecollinearSimilartothedenitionoftheepipolarlineontheimageplane,thelinecontain-ingthesethreepointsonistheintersectionoftheepipolarplanecontaining,and.Thus,thisistheepipolarlineasobservedonthereference-plane.ThecollinearityofthesethreepointscanbeexpressedasistheFundamentalMatrix.Asopposedtothereferenceframebasedformulation,wherethefundamentalmatrixdependsonthecamerarotationsandtheinternalcalibrationparametersofthecamera,hereitdependsonlyontheepipole.Moreover,theepipoleisinthehere,whereas,itisinthestandardformulation. Theexpressionfor intheP+Pcasecanberelatedtothecurrentexpressionasfollows:Consideracameracenteredat,whoseimageplaneistheplane,anditsopticalaxiscoincideswiththedirection.Then WhathappenswhentheepipolegoestoInEquation2,itcanbeseenthatwhen,theepipolegoesto.Inthiscase,andtheexpressionforparallaxcanberewrittenas:( Inotherwords,alltheparallaxvectorsareparalleltoeachother(i.e.,meetat).TheFundamentalMatrixWecan,ofcourse,unifytheniteandtheinnitecasebyusing2Dprojectivenotations.However,inthispaperwechoosetouse2DEuclideancoordinaterepresentations,inordertoemphasizethephysicalmeaningofthevariousobservedandderivedquantities.Moreover,theparallaxexpressioninEquation3,whichinvolvesmetricrelationsismeaninglessinaprojectivecoordinaterepresentation.Also,when,thengotorespectively.Thisoccurswhen,fortuitously,theplaneischosentobeparalleltotheopticrayfromthescenepointtooneofthecameras.Inthiscase,thecorrespondingimagepointcannotbeobservedonthereferenceplane,andouranalysisdoesnotapply.3DualityontheReferencePlaneInthissection,wederiveasetofdualrelationsonthereference-planebyswitch-ingtherolesofcameracentersandscenepointsaswaspreviouslydonein[2,Considertwopointsandonecameracenter.Considerthein-tersectionoftherayswiththereferenceplaneFigure2a).Theseoccurrespectivelyat.Inamanneranalogoustotheepipolarplane(denedby2cameracentersandascenepoint),wede-netheplanecontaining(2scenepointsandacameracenter)astheepipolarplane.Bythesameanalogy,wedeneitsintersectionwith(i.e.,thelineconnecting)astheepipolarline,andastheepipole.Notethatthedual-epipole,thedual-epipolarlines,andthedual-epipolarplanesrelatetoapairofscenepointsovermultipleviews,inthesamewaytheepipole,theepipolarlines,andtheepipolarplanesrelatetomultiplescenepointsoverapairofviews.Byapplyingthedualityofscenepointsandcameracenters,wecanderivethedualofthebilinearconstraintandtheparallaxexpressionsinalgebraicform.Theyare:DualParallax:),where DualBilinearityConstraint:=0, (a) ijpjt Pi PjCtpitHjHtHiPlane jtpjs Cs PjPiCtpts pitpispijPlane Fig.2.Dualityonthereferenceplane:(a)thedual-epipolargeometryassociatedwithtwopointsinoneview.istheepipole,andthebluelinegoingthroughistheepipolarline.(b)bothsetsofepipolarlines(showninred)anddual-epipolarlines(showninblue)thatarisewhenconsideringtwopointsintwoviews.is(denedas)theFundamentalMatrix.Thedualityofthebilinearconstrainthasbeenpreviouslyexplored-e.g.,Carlsson[2]andWeinshall,etal.[20]derivedualbilinearandtrilinearrelationsintermsoftheprojectivecoordinaterepresentationsofthescenepoints,cameracenters,andimagepoints.Here,however,wederivetheserelationsinthecontextofthereferenceplaneimages,andprovidephysicalmeaningtothedualrelations.Also,IraniandAnandan[9]pointedoutthedualepipoleinthecontextoftheplane+parallaxrepresentation.Inthatcase,sincetheprojectiononacameraimageplane(referenceframe)isincludedintheformulation,thereexistsanasymmetryinthevariousconstraintsandtheirdualconstraints.Here,completesymmetryisachievedbyprojectingalltheobservedquantitiesontothereferenceplaneitself.Figure2bcompletesthepicturebyconsideringtwopoints()intwoviews().Thiscongurationgivesrisetoonesetofepipolarlines(corre-spondingtoeachscenepoint)goingthroughtheepipole,andonesetofepipolarlines(correspondingtoeachcamera)goingthroughtheepipole4ThreeViewGeometryontheReferencePlaneInthissectionweextendourtreatmenttothreeviews.[5]showsthattherearenoindependentconstraintsthatcanbederivedinmorethanthreeviews.Inthissectionwepresentageometricinterpretationofthethree-viewconstraintsintermsofphysicalquantitiesonthereferenceplane.Wederivethealgebraicthree-viewconstraintsandshowthattheyhaveaverysimplemathematicalform.Wewillalsoshowthatthetensor-representationoftheseconstraintsin thereference-planehasaverysimplemathematicalformwhencomparedtothetensorsinthestandardformulations[15,7,5].4.1GeometricObservationsFigure3showsthreeviews,andofapointasprojectedontothereferenceplane.Thenewcameracenterislabeledas,andthetwonewepipolesas irpts PiCsCt pitpisCrpsrprtPlane Fig.3.Geometryofthreeviewsonthereferenceplanea3DThe3redlinesaretheepipolarlinesofpairsofviews,andtheturquoiselinethetrifocal-line .The6pointsonlieon4linesformingacompleteTakenpairwiseatatime,thethreeviewsgiverisetothreeepipolarcon-straints:arecollinear. arecollinear. arecollinear. Thereis,however,afourthcollinearityconstraint,namely:Theepipolesarecollinear Thislineissimplytheintersectionoftheplanecontainingthethreecameracenters(seeFigure3).Thisplaneisreferredtoasthetri-focalplane.Basedonthisdenitionwedenethelineconnectingthethreeepipolesasthetri-focalline.Thefactthatthesixpointslieonfourlinesisfundamentaltotheprojectionofthreeviewsofapointontoareferenceplane.Notethatthisgureonthe NotethatgeometricallythisgureisidenticaltoFigure2b,butthelabelingofthepointisdierent.ThescenepointinFigure2bhasbeenreplacedbyacameracenter.Infact,thisisbecauseofthecompletesymmetrybetweenscenepointsandcameracentersinourrepresentation. plane(Figure4a)isknownasthecompletequadrilateralandplaysacentralroleinplaneprojectivegeometry[3].Giventhethreecameras,everypointinthesceneformsatriangle(e.g.,withvertices.Dierentpoints(e.g.,indexedi,j,etc.)willformdierenttriangles,allofwhichsharethesametri-focalline(seeFigure4b).Inotherwords,allthesetrianglesareperspectivefromthetri-focalline4.2TheTrifocalRatioEachpairofviewsfromthethreeviewsprovidesanexpressionforparallaxsimilartoEquation3.Forexample,consider: (HtHi)Hs(pispts)pirpit=Hi(HtHr) Fromtheseequationswecaneliminate)toobtain: rst rst HsHr .Theaboveequationistrueuptoasignchange.Notethatrstdoesnotdependonthepoint.Inotherwords,foreveryscenepoint,thelocationsofitsimagefromthethreeviewsonthereferenceplaneisrelatedbythesameEquation5.ThisconstraintisfurtherexploredinSection4.3.Giventwoimagesofthepoint,e.g.,,andthecorre-spondingepipoles,,wecandeterminethelocationofthethirdimagebyintersectingthetwoepipolarlines(seeFigure4a).Thereare,however,twocasesinwhichthethreeepipolarlinescollapseintoasingleline(andhence,theirintersectionisnotunique).Thesearethesamesituationsnotedin[5,15],buthereweexamineitinthecontextofthereference-planeimages.Therstisthecasewhenthethreecameracentersarecollinear(seeFigure5)-inthiscasethethreeepipolescollapseintoasinglepoint(denotedinFigure5).Thethreeepipolarlinesalsocollapseintoasingleline,andcannotbedeterminedbytheintersectionoftheepipolarlines.However,giventhecommonepipolerstcanberecoveredfromusingEquation5.Infact,inthiscase,rstisthecrossratioofthesefourpoints(thethreeimagepointsandtheepipole). Itisknowninplane-projectivegeometrythatiftwotrianglesareperspectivefromalinetheyarealsoperspectivefromapoint[3] thisistheconverseoftheDesarguesTheorem.GiventhetwotrianglescorrespondingtoasinFigure4b,thenthepointofperspectivityisinfactthedual-epipole (a)ir pts pitpispsrprt ts pitpispirpsrprt pjspjrpjt Fig.4.Threeviewgeometryonthereferenceplane:(a)thecompletequadrilateralformedbytheimageofasinglepointinthreeviewsandthetrifocal-line(shownasadashedline)containingthethreeepipoles.(b)dierenttrianglesduetodierentscenepointssharethesametrifocal-line.Anotherinterestingcaseiswhenthescenepointliesonthetri-focalplaneofthethreecameras.Inthiscasethethreeimagepointsalllieonthetri-focallineitself,i.e.,onceagainthethreeepipolarlinescollapseontothetri-focalline.Hencewecannotusetheintersectionofepipolarlinesto.Inthiscasetoo,canbedeterminedbyEquation5,usingrstTheratiorsthasaspecialsignicance.Ifweconsiderthetri-focalline,wecanshow(byreplacinginEquation3)that:rst)(6)(Hence,thenametrifocal-ratio.)Inotherwords,inthegeneralcase:rst ||psrpts||=||pispit|| ||pispts||||pirp Notethatinthesingularcase,whentheepipolescollapse,theratioofthedis-tancesbetweentheepipoles(thetopequation)isundened,butthebottomequationisstillvalidandcanbeused.4.3TheTrifocalTensorReturningtoEquation5,wecanwritedowncomponentequalitiesasfollows: xisxts=yisyit rst rst Bytakingtwooftheseequalitiesatatimeandcross-multiplyingbythedenomi-natorswecangetsixlinearconstraints.Ofthesetwoarethesameasthebilinear(epipolar)constraintsinvolvingonlytwoviewsatatime.Theotherfour,whichinvolveallthreeviewsare:rstrstrstrst)(9) pite = pts = psr = prtpirpis Fig.5.TheEpipolarLinesCollapseNotethatthesethreeviewconstraintsareactuallyonlybilinearintheimagelocationsofthescenepoint(asopposedtothetrilinearconstraintsin[15])).Thisisbecausebyconsideringtheprojectionofthepointsonthereferenceplaneitself,weeliminatethehomographiesinducedbetweentheviews(whichappearin[16]).ThetrilinearformsgiveninEquation9canbeuniedintoasingletensorequationinamanneranalogousto[17]:rst)=0(10),2indicatetherowindicesofof10xis])4.Basedonfurtheralgebraicmanipulation,Equation10canberewrittenas:,20=rstrstfollowsthestandarddenition:=1ifand0otherwise.rstis33tensorrst=((rstIntheaboveequations,(,etc.denotetherst(i.e.,),thesecond(i.e.,),andthethird(i.e,1)componentsof,etc.Similarly(rstdenotestheentryindexedbya,b,cintheTensor. Notethatasin[17],istheverticallineonpassingthroughisthehorizontallineonpassingthrough.Similarlyaretheverticalandhorizontallinesonpassingthrough.Also,asin[17]therelationshipsinEqua-tion10arevalidforanylinepassingthroughandanyotherlinepassingthrough.Inotherwords,Equation10capturesthesamepoint-line-linerelationshipde-scribedin[17]and[5]. Notethattheelementsofrstdependonthetwoepipolesrst.Thisisincontrasttothegeneralformofthetrifocaltensor forexample,thetrilineartensorin[15]alsodependsonthehomographiesduetotheplanebetweenthedierentcamerasandthetensordescribedin[5]whichdependsonthecameraprojectionmatrices.AsinthecaseoftheFundamentalMatrixinourformulation,theepipolesareexplicitwithintheTensor,whereasinthegeneralformulation,thetensorisimplicitlyrelatedtotheepipole.GiventheTensorrstwecanrecoverthetwoepipolesandthetrifocal-rationrst;usingEquation6wecanrecoverthethirdepipole4.4DualsoftheThreeViewConstraints3ScenePoints+1Camera:Asinthecaseoftwo-viewanalysis,thedualitybetweenscenepointsandcameracentersalsoappliestothree-viewanalysis.ByswitchingtherolesofscenepointsandcameracentersinFigure3(i.e.,)wecanderivenewconstraintsinvolvingonecameracenterandthreepoints.Theresultinggeometriccongurationisalsocompletequadrilateral,butwithadierentlabelingofthepoints.Figure6aindicatesthelabelingcorrespondingtooneviewofthreepoints.Inthiscasetrifocal-linecontainsthedual-epipoles,and.Thethree-viewconstraintgiveninEquation5isreplacedby ijk ijk HjHk ,isthetothetrifocal-ratiorst.Dualtotheotherformsofthethree-viewconstraints,(e.g.,Equation9)canalsobeobtainedbythesamesubstitutionofindices(i.e,t,ti,sj,rleadingtothe-tensorform:ijk=((ijkandthecorrespondingconstraintset:,20=ijkwherethedenitionsofthe23matricesareanalogoustothedenitionsgivenearlier.Notethatthethethree-viewconstraintsandtheirdual,thethree-pointconstraintsarecompletelysymmetric.Thedual-tensordependsonthedual-epipolesandthetothetrifocal-ratioOtherCombinationsof3+1Points:Thecompletesymmetrybetweenscenepointsandcameracentersimpliesthatwecanarbitrarilychoosethelabel(eitherasascenepointorasacameracenter)foreachofthefour3DpointsinFig-ure3.Sofar,wehaveconsideredtwochoices:3cameracenter+1scenepoint , (a)ki pjk pjtpktpitpij itpispjtpjs pijpts Fig.6.DualstotheThreeViewGeometry:(a)thecompletequadri-lateralformedby3points+1cameracenter .(b)thequadrilateralformedby2points+2cameras .Notethattheepipolar-lines(thinlines)intersectatanepipole,thedual-epipolarlines(thicklines)intersectatadual-epipole,andanepipolarlineintersectsadual-epipolarlineatanimage-pointand3scenepoints+1cameracenter .Thebasicstructureisthatfourpointsaredividedintoagroupof3andasinglepoint.Wecanobtainotherdualsbychoosingthefourpointstoconsistof2cameracenters+2scenepoints groupingthemas2cameracentersandascenepoint+1scenepoint oras2scenepointsandacameracenter+1cameracenter InFigure6bweshowtheresultingquadrilateralcorrespondingtotherstofthesegroupings.Sincethecongurationshowninthisgureisbasedon2cameracentersand2scenepoints,thesixpointsonthequadrilateralconsistoffourimagepoints,oneepipole,andonedual-epipole.Notethatthetwoepipolarlinesintersectatanepipole,thetwodual-epipolarlinesintersectatadual-epipoleandeachepipolarlineintersectseachdual-epipolarlineatanimagepoint.Unlikethe3Dworld,wheretherearetwotypesofpoints,cameracentersandscenepoints,onthereference-plane,therearethree-typesofpoints epipoles,dual-epipoles,andimagepoints.Eachoftheseformthecenterofaradialeldoflinesthatgothroughthatpoint,allthreehavecompletelydual-roleson5ApplicationsThesimpleandintuitiveexpressionsderivedinthereference-planebasedformu-lationinthispaperleadtousefulapplicationsin3Dsceneanalysis.Inparticular,thesimpler(bilinear)tri-focalconstraintswiththeidentiedtri-focalratioleadtoasimplemethodforNewViewSynthesis.Initialexperimentalresultsareshowninthissection.Moreover,theseparationandcompactpackingofthecameracalibrationandorientationintothe2Dprojectiontransformation(ahomography)thatrelatestheimageplanetothereferenceplane,leadstopotentiallypowerfulreconstructionandcalibrationalgorithms.Thesearebrieydiscussedinthissection.5.1NewViewGenerationUsingtheThree-ViewConstraintsInthissectionweshowthatthereference-planebasedformulationprovidesasimpleandintuitivewaytogeneratenewviewsfromagivensetofviews.Werstshowsomeresults,followedbyanexplanationhowtheywereob- Figure7aand7bdisplaytwoimagestakenbyahand-heldcamera.Thescenecontainedtoyswhichwereplacedonarugontheoor.Thecameratranslatedandrotatedbetweenthetwoviews.The3Dparallaxeectsduetothecameratranslationareapparentinthechangeofthe2Ddistance(ontheimage)betweentheclownshatandtheupper-rightcorneroftherug.Figure7cisanewsynthesizedviewofthescene,asifobtainedfromacamerapositionedfarthertotheleftofthescenerelativetothetwooriginalviews(androtated,tocompensateforthetranslation).Notethesmallerdistancebetweentheclownshatandthecorneroftherug.Forcomparisonandverica-tion,Figure7dshowsanviewobtainedfromthesameviewingdirectionandorientation.Also,notethedierencesbetweentheactualandsynthesizedview.Thereareimagedistortionswheretheowwasinaccurate(atdepthdis-continuities,e.g.,ontherugaroundtheclownshead,andneartheearsofthesmallerdoll).Also,thesynthesizedviewismissingtheleftpartoftherug,asthisportionoftherugwasnotviewedinanyofthe2inputimages. (b) (d) (c) Fig.7.NewViewSynthesis.(a)and(b)showtwoimagestakenbyahand-heldcamera.Thecameratranslatedandrotatedbetweenthetwoviews.The3Dparallaxeectsduetothecameratranslationareapparentbythechangeinthe2Ddistance(ontheimage)betweentheclownshatandtheupper-rightcorneroftherug.(c)Anewsynthesizedviewofthescene,Notethesmallerdistancebetweentheclownshatandthecorneroftherug.(d)anviewobtainedfromthesameviewingdirectionandorientation.Notethedierencesbetweentheactualandsynthesizedview:Thereareimagedistortionswheretheowwasinaccurate(e.g.,ontherugaroundtheclownshead,andneartheearsofthesmallerdoll).Also,thesynthesizedviewismissingtheleftpartoftherug,asthisportionoftherugwasnotviewedinanyofthe2inputimages,(aandb). Belowisabriefdescriptionofhowthesynthesizedviewwasgenerated.Toworkdirectlywithquantitiesonthereferenceplanewouldrequirepartialcal-ibrationinformationabouttheinputviews.Butasexplainedbelow,newviewsynthesisispossibleevenwithoutsuchinformation.StepI:Oneofthetwoinputimages(camera)isrstwarpedtowardstheotherinputimage(camera;thereferenceimage)viaa2Dprojectivetrans-formationtoaligntheoftheplaneinthetwoinputimageistheplaneoftherug,inourcase).Thecorresponding2Dprojectivetransfor-mationiscomputedautomatically,withoutanypriororadditionalinformation,usinga2Dregistrationtechniquedescribedin[10].Thismethodlocksontoa2Dparametrictransformationbetweenapairofimages,eveninthepresenceofmovingobjectsorotheroutliers(suchasthetoys,inourcase).Formoredetailssee[10].Notethataftersuch2Dwarping,thetwoplane-stabilizedimagesareinfullalignmentinallimageregionswhichcorrespondtotherug,andaremisalignedinallother(i.e.,out-of-plane)imagepoints(i.e.,thetoys).Thefartherascenepointisfromtheplanarsurface(rug),thelargeritsresidualmisalignment.Werefertotheseasplanar-parallaxdisplacements(see[11,12,9]).Notethattheplane-stabilizedsequenceisinfacta2Dre-projectionofthecorrespondingvirtualimagesonthereferenceplaneontothereferenceim-ageplane,(SeeFigure1.b).Therefore,aquadrilateralonwillprojecttoaquadrilateralontheimageplane;dierenttrianglesoncorrespondingtodierentscenepointsandsharingacommontri-focallinewillpreservethisrela-tiononthereferenceimageplane.Itcanbeshownthatforanyquadrilateral,thereexistssomerstsuchthatEquation(7)holds.Infact,itcanbeshownrst HsHr isthecomponentofthetranslationbetweencamerastheoptical()axisofthereferencecamera.Similarlyforthethirdcameraareasbefore(i.e.,heightsrelativetoStepII:Denseowisestimatedbetweenthetwoplane-stabilizedimages(us-ingthemethoddescribedin[12]).Notethatafterplanestabilization,theoweldbetweenthetwoimagesreducestoaradialepipolareldcenteredattheepipole(seeEquation(3);seealso[11,12,9]).Thecancellationoftheplaneho-mographyremovesalleectsofcamerarotationandchangesincalibration.Thisallowstocomputetheoweldbetweenaplane-stabilizedimagepairmorereli-ablythangeneralow,asitisconstrainedtosatisfyaglobalepipolarconstraint.StepIII:Weestimatetheepipole()fromtheradialoweldbetweenthetwoinputplane-stabilizedimages.Wethenspecify:(i)theepipole(e.g.,)betweenthereferenceimageandtheplane-stabilizedimage,(ii)atri-focalratiorstinthe referenceframe.Giventhevirtualtri-focalratiorst,thevirtualepipoletheactualepipole,andthedenseoweldbetweenthetwoplane-stabilizedimages(betweensandcorrespondings),wecanestimateallimagepointsinthevirtual(plane-stabilized)image(namely,alls)usingEquation8.Thevirtualtri-focalratiorstandthevirtualepipolecaneitherbespec-ieddirectly(e.g.,viaEquation(14)),orelsebyspecifyingthelocationoftwoormoreimagepointsinthevirtual(plane-stabilized)view,andestimatethemaccordingly.StepIV:Notethatthesynthesizedplane-stabilizedimageisthesameforcameracenteredat.Inotherwords,itisindependentoftheinternalparame-tersandtheorientationofthatcamera.Byspecifyingahomographythatrelatestheimageplaneofthevirtualcameratothestabilizedimagefromthereferenceview,wehavethecompleteexibilitytogenerateanimageobtainedbyerasituatedat.Thisisdonebythesynthesizedplane-stabilizedimageviathecorresponding2Dprojectivetransformation.5.23DReconstructionandCameraCalibrationGivenuncalibratedimages,anyapproachforobtainingEuclidean(orAne)reconstructionrequiressometypeofcalibration.Oneofthebenetsofourap-proachisthatthisprocessisfactoredintotwoseparatestages,eachofwhichhasasimpleandintuitivesolution.First,giventheinputimages,thevirtual-imagesonmustbedetermined.ThiscanbedonebytakingadvantageoftheP+Pmethod[12] (i)determinetheplanarhomographyforbetweenarbitrarilychosenreferenceimageandeachotherimage,and(ii)determinethehomographybetweenthereferenceimageand.NotethattheparallaxEqua-tion3isvalideveniftheimagelocationsonareknownonlyuptoa2Dtransformation.Thismeansthatjustbyindicatingtwosetsofparallellines(thatareindierentorientations),the3DHeightsrelativetothereferenceplanecanberecovered.Fromthisinformation,theparallaxmagnitude(inEqua-tion3)canbedetermined.(Notethatbyspecifyingthe2Dcoordinatesoffourpointsonthereferenceplane,thehomographycanfullydetermined,leadingtoEuclideanreconstruction.)Givenandtheheightofone3DscenepointrelativetotheHeightsofallotherpointscanbedetermineduptoaglobalscalefactor.Boththesecalibrationstepsaresimpleandintuitiveandrequireminimalspecicationofinformation.Theresultingreconstructioniswithrespecttothereferenceplaneanddoesnotinvolvethecamerareferenceframes.Theforegoingoutlineforareconstructionmethodassumesthatthecor-respondencesofeachpointacrossthemultiple-viewscanbeestimated.Thisinvolvescomputingtheparallaxow-eld(s),andtheepipole(s) thesecanbedoneinthesamemannerasdescribedin[11,12].Itisworthnoting,however,theremovaloftheplanarhomographyallowstheparallaxcomputationtobemorerobustandaccurate[11,12]. 1.S.AvidanandA.Shashua.Novelviewsynthesisintensorspace.InIEEECon-ferenceonComputerVisionandPatternRecognition,pages1034 1040,San-Juan,June1997.2.S.Carlsson.Dualityofreconstructionandpositioningfromprojectiveviews.InWorkshoponRepresentationsofVisualScenes,1995.3.H.S.MCoxeter,editor.ProjectiveGeometry.SpringerVerlag,1987.4.O.D.Faugeras.Whatcanbeseeninthreedimensionswithanuncalibratedstereorig?InEuropeanConferenceonComputerVision,pages563 578,SantaMargaritaLigure,May1992.5.O.D.FaugerasandB.Mourrain.Onthegeometryandalgebraofthepointandlinecorrespondencesbetweennimages.InInternationalConferenceonComputer,pages951 956,Cambridge,MA,June1995.6.OlivierFaugeras.Three-DimensionalComputerVision AGeometricViewpointMITPress,Cambridge,MA,1996.7.RichardHartley.Linesandpoinsinthreeviews auniedapproach.InDARPAImageUnderstandingWorkshopP,1994.8.RichardHartley.EuclideanReconstructionfromUncalibratedViews.InApplica-tionsofInvarianceinComputerVision,J.L.Mundy,D.Forsyth,andA.Zisserman(Eds.),Springer-Verlag,1993.9.M.IraniandP.Anandan.Parallaxgeometryofpairsofpointsfor3dsceneanalysis.EuropeanConferenceonComputerVision,Cambridge,UK,April1996.10.M.Irani,B.Rousso,andS.Peleg.Computingoccludingandtransparentmotions.InternationalJournalofComputerVision,12(1):5 16,January1994.11.M.Irani,B.Rousso,andP.peleg.Recoveryofego-motionusingregionalignment.IEEETrans.onPatternAnalysisandMachineIntelligence,19(3):268 272,March12.R.Kumar,P.Anandan,andK.Hanna.Directrecoveryofshapefrommultipleviews:aparallaxbasedapproach.InProc12thICPR,1994.13.H.C.Longuet-Higgins.Acomputeralgorithmforreconstructingascenefromtwoprojections.Nature,293:133 135,1981.14.R.Mohr.AccurateProjectiveReconstructionInApplicationsofInvarianceinComputerVision,J.L.Mundy,D.Forsyth,andA.Zisserman,(Eds.),Springer-Verlag,1993.15.A.Shashua.Algebraicfunctionsforrecognition.IEEETransactionsonPatternAnalysisandMachineIntelligence,17:779 789,1995.16.A.ShashuaandN.Navab.Relativeanestructure:Theoryandapplicationto3dreconstructionfromperspectiveviews.InIEEEConferenceonComputerVisionandPatternRecognition,pages483 489,Seattle,Wa.,June1994.17.A.ShashuaandP.AnanadanTrilinearConstraintsrevisited:generalizedtrilinearconstraintsandthetensorbrightnessconstraint.IUW,Feb.1996.18.P.H.S.Torr.MotionSegmentationandOutlierDetection.PhDThesis:ReportNo.OUEL1987/93,Univ.ofOxford,UK,1993.19.M.SpetsakisandJ.Aloimonos.Auniedtheoryofstructurefrommotion.DARPAImageUnderstandingWorkshop,pp.271-283,Pittsburgh,PA,1990.20.D.Weinshall,M.Werman,andA.Shashua.Shapedescriptors:Bilinear,trilinearandquadlinearrelationsformulti-pointgeometry,andlinearprojectivereconstruc-tionalgorithms.InWorkshoponRepresentationsofVisualScenes,1995.