计算机视觉computer vision 11.docx
- 文档编号:12563211
- 上传时间:2023-06-06
- 格式:DOCX
- 页数:14
- 大小:297.55KB
计算机视觉computer vision 11.docx
《计算机视觉computer vision 11.docx》由会员分享,可在线阅读,更多相关《计算机视觉computer vision 11.docx(14页珍藏版)》请在冰点文库上搜索。
计算机视觉computervision11
Chapter11
ShapefromPhotographs:
AMulti-viewStereoPipeline
Abstract.
Acquiring3DshapefromimagesisaclassicprobleminComputerVisionoccupyingresearchersforatleast20years.Onlyrecentlyhoweverhavethese
ideasmaturedenoughtoprovidehighlyaccurateresults.Wepresentacompletealgorithmtoreconstruct3Dobjectsfromimagesusingthestereocorrespondencecue.
Thetechniquecanbedescribedasapipelineoffourbasicbuildingblocks:
camera
calibration,imagesegmentation,photo-consistencyestimationfromimages,and
surfaceextractionfromphoto-consistency.InthisChapterwewillputmoreemphasisonthelattertwo:
namelyhowtoextractgeometricinformationfromasetof
photographswithoutexplicitcameravisibility,andhowtocombinedifferentgeometryestimatesinanoptimalway.
11.1Introduction
Digitalmodelingof3Dscenesisbecomingincreasinglypopularandnecessaryforawiderangeofapplicationssuchasculturalheritagepreservation,onlineshoppingorcomputergames.Althoughactivemethodsremainoneofthemostpopulartechniquesofacquiringshape,thehighcostoftheequipment,complexity,anddifficultiestocapturecolorarethreebigdisadvantages.Asopposedtoactivetechniques,photograph-basedtechniquesprovideanefficientandeasywaytoacquire
shapeandcolorbysimplycapturingasequenceofphotographsoftheobject.Thegoalofanyshape-from-photographsalgorithmcanbedescribedas“givenasetofinputphotographs,howtoestimatea3Dshapethatwouldgeneratethesamephotographs,assumingsamematerial,viewpointsandlightingconditions”.
Thisdefinitionhighlightsthemaindifficultyoftheproblem:
photographsareobtainedasaresultofcomplexinteractionsbetweenthegeometryofthescene,thematerialsofthescene,thelightingconditionsandtheviewpoints(seeFigure11.1).
Hencerecoveringthegeometryjustfromphotographsisnotonlyachallengingproblembutalso,inthegeneralcase,anill-posedproblem.Itischallengingbecauselightingandmaterialpropertiesplayaveryimportantroleintheimageformationmodel.Thesamegeometrywithadifferentmaterialordifferentlightingconditionscangiveextremelydifferentphotographs.Itisalsoanill-posedproblembecause,inthegeneralcase,differentcombinationsofgeometry,lightingandmaterialcanproduceexactlythesamephotographs,makingitimpossibletorecoverasinglescenegeometry.Themainrecipetomaketheproblemwell-posedistousepriorsonthetypesofsurfacethatoneexpects.Traditionallythemostcommontypeofprioristhesmoothsurfaceprior.Howeverwhendealingwithspecialclassesofobjectssuchashumanfacesorman-madeobjects,moreevolvedpriorshavebeensuccessfullyused(e.g.,humanfaces,buildingsorplanes).
Asfortheimportanceofmaterialsandlightingconditions,ithasbeenaddressed
byrestrictingtheclassofmaterialsaparticularalgorithmisdesignedfor.Asaresult,nosinglemethodisabletocorrectlyreconstructageneralscenewithanytypeofmaterialsandlightingconditions,leadingtoaplethoraofspecificalgorithmsdesignedforspecifictypesofobjectsandusingspecificcues:
silhouettes,texture,transparency,defocus,shadingorcorrespondence,bothsparseanddense.Historicallythemostsuccessfulcueshavebeensilhouettes,correspondence,andshading.Silhouettesandcorrespondencesarethemostrobustofallduetotheirinvariancetoilluminationchanges.Theshadingcueneedsamorecontrolledilluminationenvironment,butitcanproducebreathtakingresults,whichmakesitwidelyusedtoo.AnexampleofanalgorithmexploitingtheshadingcueisshowninFigure11.2.Thealgorithmisdesignedtofinda3Dshapethatproducesthesameshadingastheoriginalobject.Interestingly,iftheestimate3Dshapeisthenusedtomanufactureareplicafromadifferentmaterial(inFigure11.2theoriginalisporcelain,whilethereplicaisplaster)wecanappreciatehowthereplicastillshowsthesameshadingpattern.Thisisthedesiredbehavior,sincethealgorithmisspecificallydesignedtoimitatetheshading,nottoproduceidenticalphotographies.
Amongthevastliteratureavailableonimage-basedmodelingtechniques,recent
workonmulti-viewstereo(MVS)reconstructionhasbecomeagrowingareaofinterestinrecentyearswithmanydifferingtechniquesachievingahighdegreeof
accuracy.Thesetechniquesaremainlybasedonthecorrespondencecueand
focusonproducing3Dmodelsfromasequenceofcalibratedimagesofanobject,
wheretheintrinsicparametersandposeofthecameraareknown.Inadditiontoprovidingataxonomyofmethods,alsoprovidesaquantitativeanalysisofperformancebothintermsofaccuracyandcompleteness.Ifwetakealookatthetopperformers,theymaybelooselydividedintotwogroups.Thefirstgroupmakesuseoftechniquessuchascorrespondenceestimation,localregiongrowingandfilteringtobuildupa“cloudofpatches”thatcanbeoptionallymadedenseusingmeshingalgorithmssuchasPoissonreconstructionorsigneddistancefunctions.Thesecondgroupmakesuseofsomeformofglobaloptimizationstrategyonavolumetricrepresentationtoextractasurface.Underthissecondparadigm,a3Dcostvolumeiscomputed,andthena3Dsurfaceisextractedusingtoolspreviouslydevelopedforthe3Dsegmentationproblemsuchasdeformablemodels,level-setsorgraph-cuts.
Thewayvolumetricmethodsusuallyexploitphoto-consistencyisbybuildinga3Dmapofphoto-consistencywhereeach3Dlocationgivesanestimateofhowphoto-consistentwouldbethereconstructedsurfaceatthatlocation.Theonlyrequirementtocomputethisphoto-consistency3Dmapisthatcameravisibilityisavailable.Unfortunately,thegeometryofthescene,i.e.,whatwetrytocompute,isrequiredtoknowwhichcamerasseea3Dlocation(seeFigure11.3).
Inordertobreakthisdependencybetweenvisibilityandshape,multi-viewstereoalgorithmshavetakendifferentapproaches.Amajorityofmethodsusethenotionof“currentsurface”inordertojointlyoptimizeforcameravisibilityandshape.Thevisibilitycomputedfromthereconstructedsurfaceatiterationi−1isthenusedtocomputephoto-consistencyatiterationi,improvingthereconstructiongradually.InthisChapterwewillgivefurtherinsightintoatwo-stageMVSvolumetricapproach:
namelyhowtoextracta3Dvolumeofphoto-consistencyfromasetofphotographswithoutexplicitcameravisibilityinSection11.3,andhowtoextractasurfacefromthephoto-consistencyvolumeinagloballyoptimalwayinSection11.4.ThepipelinedescribedinthisChapteriscurrentlyatopperformerintherecentevaluationofmulti-viewstereoalgorithmsbySeitzetal.
11.2Multi-viewStereoPipeline:
FromPhotographsto3DModels
Thereexistsavastliteratureonmulti-viewstereoalgorithms.Eventhoughmany
ofthemethodssharethesamebasicarchitecture,theydiffermainlyinwhattype
ofscenesorcomputationtimetheyareoptimizedtoworkwith.Allthemulti-view
stereomethodsusethecorrespondencecue,whichisusuallyexploitedintheform
ofaphoto-consistencymetricsuchasNormalizedCrossCorrelation,SumofSquare
Differences,orMutualInformation.Startingfromthephoto-consistencymetric,differentalgorithmsfocusondifferenttargetapplicationssuchasoutdoorscenes,buildingreconstruction,interiorbuildingsorobjectreconstruction.InthisChapterwedescribeavolumetricmulti-viewstereoapproachthatisoptimizedforgeneralscenereconstruction,withapreferenceforwatertightsurfaces.Thepipeline(seeFigure11.4)canbedescribedas:
•photographacquisition,
•cameracalibration,
•computing3Dphoto-consistencyfromasetofcalibratedphotographs,
•extractinga3Dsurfacefroma3Dmapofphoto-consistency.
InthefollowingSectionswefocusonhowtoextract3Dphoto-consistencyfromasetofphotographs(seeSection11.3)andhowtousethe3Dphoto-consistencytoextracta3Dsurface(seeSection11.4).Weleavethediscussiononimageacquisition,e.g.,real-timevsphotograph-based,andoncameracalibrationforfuturediscussion(seeforanstate-of-the-artsystemtocalibrateasetofphotographs).
11.3ComputingPhoto-ConsistencyfromaSetofCalibratedPhotographs
Givenasetofimagesandtheircorrespondingcameraposes,wewouldliketoextracta3Dmapofphoto-consistencythattellushowphoto-consistentisaparticular3Dlocationforagivensetofvisiblecameras.Themaindifficultyofthisstepishowtoproduceavolumetricmeasureofphoto-consistencywithouttheknowledgeofthesetofcamerasthatshouldbeusedtocomputephoto-consistencyforevery3Dlocation.
Thisproblemisaddressedintheproposed3Dmodelingpipelinebyfollowingasimilarapproachtowherephoto-consistencyismaderobusttoocclusion.Thisapproachcomputesa3Dmapofphoto-consistencyasanaggregationofdepth-mapsfromdifferentview-points(seeFigure11.5).Thecreationofsuchaphoto-consistency3Dmapissimilarinspirittothespacecarvingapproachproposedby.However,bycomputingitasanaggregationofdepth-maps,twoadvantagesappear:
•depth-mapcomputationusingdensestereoisaverysuccessfulandactiveresearchtopic.Itisanidealbuildingblocktousesinceimprovementsinthefieldofdensestereocanbedirectlybeneficialtothemulti-viewstereoproblem.
•Computationtimeisnolongerdependentontheresolutionofthe3Dvolume,butonthenumberofcameras.Itisalsohighlyparallelizable,sinceeachdepth-mapisindependentlycomputedandnoiteratedvisibilitycomputationisrequired.
Bybuildinga3Dmapofphoto-consistency,the3Dreconstructionproblemcannowbeseenasa3
- 配套讲稿:
如PPT文件的首页显示word图标,表示该PPT已包含配套word讲稿。双击word图标可打开word文档。
- 特殊限制:
部分文档作品中含有的国旗、国徽等图片,仅作为作品整体效果示例展示,禁止商用。设计者仅对作品中独创性部分享有著作权。
- 关 键 词:
- 计算机视觉computer vision 11 计算机 视觉 computer