




版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
FeaturesDigitalVisualEffectsYung-YuChuangwithslidesbyTrevorDarrell
CordeliaSchmid,DavidLowe,DaryaFrolova,DenisSimakov,RobertCollinsandJiwonKimOutlineFeaturesHarriscornerdetectorSIFTExtensionsApplicationsFeaturesFeaturesAlsoknownasinterestingpoints,salientpointsorkeypoints.Pointsthatyoucaneasilypointouttheircorrespondencesinmultipleimagesusingonlylocalinformation.?DesiredpropertiesforfeaturesDistinctive:asinglefeaturecanbecorrectlymatchedwithhighprobability.Invariant:invarianttoscale,rotation,affine,illuminationandnoiseforrobustmatchingacrossasubstantialrangeofaffinedistortion,viewpointchangeandsoon.Thatis,itisrepeatable.ApplicationsObjectorscenerecognitionStructurefrommotionStereoMotiontracking…ComponentsFeaturedetectionlocateswheretheyareFeaturedescriptiondescribeswhattheyareFeaturematchingdecideswhethertwoarethesameoneHarriscornerdetectorMoraveccornerdetector(1980)WeshouldeasilyrecognizethepointbylookingthroughasmallwindowShiftingawindowinany
directionshouldgivealargechangeinintensityMoraveccornerdetectorflatMoraveccornerdetectorflatMoraveccornerdetectorflatedgeMoraveccornerdetectorflatedgecornerisolatedpointMoraveccornerdetectorChangeofintensityfortheshift[u,v]:windowfunctionFourshifts:(u,v)=(1,0),(1,1),(0,1),(-1,1)Lookforlocalmaximainmin{E}intensityshiftedintensityProblemsofMoravecdetectorNoisyresponseduetoabinarywindowfunctionOnlyasetofshiftsatevery45degreeisconsideredOnlyminimumofEistakenintoaccountHarriscornerdetector(1988)solvestheseproblems.HarriscornerdetectorNoisyresponseduetoabinarywindowfunctionUseaGaussianfunction
HarriscornerdetectorOnlyasetofshiftsatevery45degreeisconsideredConsiderallsmallshiftsbyTaylor’sexpansionHarriscornerdetectorOnlyasetofshiftsatevery45degreeisconsideredConsiderallsmallshiftsbyTaylor’sexpansionHarriscornerdetectorEquivalently,forsmallshifts[u,v]wehaveabilinearapproximation:,whereMisa22matrixcomputedfromimagederivatives:Harriscornerdetector(matrixform)HarriscornerdetectorOnlyminimumofEistakenintoaccountAnewcornermeasurementbyinvestigatingtheshapeoftheerrorfunctionrepresentsaquadraticfunction;Thus,wecananalyzeE’sshapebylookingatthepropertyofMHarriscornerdetectorHigh-levelidea:whatshapeoftheerrorfunctionwillwepreferforfeatures?flatedgecornerQuadraticformsQuadraticform(homogeneouspolynomialofdegreetwo)ofnvariablesxiExamples=SymmetricmatricesQuadraticformscanberepresentedbyarealsymmetricmatrixAwhereEigenvaluesofsymmetricmatricesBradOsgoodEigenvectorsofsymmetricmatricesEigenvectorsofsymmetricmatricesHarriscornerdetectorIntensitychangeinshiftingwindow:eigenvalueanalysis
1,
2–eigenvaluesofMdirectionoftheslowestchangedirectionofthefastestchange(
max)-1/2(
min)-1/2EllipseE(u,v)=constVisualizequadraticfunctionsVisualizequadraticfunctionsVisualizequadraticfunctionsVisualizequadraticfunctionsHarriscornerdetector
1
2Corner
1and
2arelarge,
1~
2;
Eincreasesinalldirections
1and
2aresmall;
Eisalmostconstantinalldirectionsedge
1>>
2edge
2>>
1flatClassificationofimagepointsusingeigenvaluesofM:HarriscornerdetectorMeasureofcornerresponse:(k–empiricalconstant,k=0.04-0.06)Onlyforreference,youdonotneedthemtocomputeRHarriscornerdetectorAnotherviewAnotherviewAnotherviewSummaryofHarrisdetectorComputexandyderivativesofimageComputeproductsofderivativesateverypixelComputethesumsoftheproductsofderivativesateachpixelSummaryofHarrisdetectorDefinethematrixateachpixelComputetheresponseofthedetectorateachpixelThresholdonvalueofR;computenonmaxsuppression.Harriscornerdetector(input)CornerresponseRThresholdonRLocalmaximumofRHarriscornerdetectorHarrisdetector:summaryAverageintensitychangeindirection[u,v]canbeexpressedasabilinearform:
DescribeapointintermsofeigenvaluesofM:
measureofcornerresponse
Agood(corner)pointshouldhavealargeintensitychangeinalldirections,i.e.RshouldbelargepositiveNowweknowwherefeaturesareBut,howtomatchthem?Whatisthedescriptorforafeature?Thesimplestsolutionistheintensitiesofitsspatialneighbors.Thismightnotberobusttobrightnesschangeorsmallshift/rotation.()123456789123456789Harrisdetector:somepropertiesPartialinvariancetoaffineintensitychange
Onlyderivativesareused=>invariancetointensityshiftI
I
+
b
Intensityscale:I
a
IRx
(imagecoordinate)thresholdRx
(imagecoordinate)HarrisDetector:SomePropertiesRotationinvarianceEllipserotatesbutitsshape(i.e.eigenvalues)remainsthesameCornerresponseRisinvarianttoimagerotationHarrisDetectorisrotationinvariantRepeatabilityrate:#correspondences
#possiblecorrespondencesHarrisDetector:SomePropertiesBut:notinvarianttoimagescale!AllpointswillbeclassifiedasedgesCorner!Harrisdetector:somepropertiesQualityofHarrisdetectorfordifferentscalechangesRepeatabilityrate:#correspondences
#possiblecorrespondencesScaleinvariantdetectionConsiderregions(e.g.circles)ofdifferentsizesaroundapointRegionsofcorrespondingsizeswilllookthesameinbothimagesScaleinvariantdetectionTheproblem:howdowechoosecorrespondingcirclesindependentlyineachimage?ApertureproblemSIFT
(ScaleInvariantFeatureTransform)SIFTSIFTisancarefullydesignedprocedurewithempiricallydeterminedparametersfortheinvariantanddistinctivefeatures.SIFTstages:Scale-spaceextremadetectionKeypointlocalizationOrientationassignmentKeypointdescriptor()localdescriptordetectordescriptorA500x500imagegivesabout2000features1.Detectionofscale-spaceextremaForscaleinvariance,searchforstablefeaturesacrossallpossiblescalesusingacontinuousfunctionofscale,scalespace.SIFTusesDoGfilterforscalespacebecauseitisefficientandasstableasscale-normalizedLaplacianofGaussian.DoGfilteringConvolutionwithavariable-scaleGaussianDifference-of-Gaussian(DoG)filterConvolutionwiththeDoGfilterScalespace
doublesforthenextoctaveK=2(1/s)Dividingintooctaveisforefficiencyonly.Detectionofscale-spaceextremaKeypointlocalizationXisselectedifitislargerorsmallerthanall26neighborsDecidescalesamplingfrequencyItisimpossibletosamplethewholespace,tradeoffefficiencywithcompleteness.Decidethebestsamplingfrequencybyexperimentingon32realimagesubjecttosynthetictransformations.(rotation,scaling,affinestretch,brightnessandcontrastchange,addingnoise…)DecidescalesamplingfrequencyDecidescalesamplingfrequencys=3isthebest,forlargers,toomanyunstablefeaturesfordetector,repeatabilityfordescriptor,distinctivenessPre-smoothing=1.6,plusadoubleexpansionScaleinvariance2.AccuratekeypointlocalizationRejectpointswithlowcontrast(flat)andpoorlylocalizedalonganedge(edge)Fita3Dquadraticfunctionforsub-pixelmaxima1650-1+12.AccuratekeypointlocalizationRejectpointswithlowcontrast(flat)andpoorlylocalizedalonganedge(edge)Fita3Dquadraticfunctionforsub-pixelmaxima1650-1+12.AccuratekeypointlocalizationTaylorseriesofseveralvariablesTwovariablesAccuratekeypointlocalizationTaylorexpansioninamatrixform,xisavector,fmapsxtoascalarHessianmatrix(oftensymmetric)gradient2Dillustration2Dexample-17-1-17777-9-9DerivationofmatrixformDerivationofmatrixformDerivationofmatrixformDerivationofmatrixformDerivationofmatrixformAccuratekeypointlocalizationxisa3-vectorChangesamplepointifoffsetislargerthan0.5Throwoutlowcontrast(<0.03)AccuratekeypointlocalizationThrowoutlowcontrastEliminatingedgeresponsesr=10LetKeepthepointswithHessianmatrixatkeypointlocationMaximainDRemovelowcontrastandedgesKeypointdetector233x89832extrema729aftercon-trastfiltering536aftercur-vaturefiltering3.OrientationassignmentByassigningaconsistentorientation,thekeypointdescriptorcanbeorientationinvariant.Forakeypoint,ListheGaussian-smoothedimagewiththeclosestscale,orientationhistogram(36bins)(Lx,Ly)mθOrientationassignmentOrientationassignmentOrientationassignmentOrientationassignmentσ=1.5*scaleofthekeypointOrientationassignmentOrientationassignmentOrientationassignmentaccuratepeakpositionisdeterminedbyfittingOrientationassignment36-binorientationhistogramover360°,weightedbymand1.5*scalefalloffPeakistheorientationLocalpeakwithin80%createsmultipleorientationsAbout15%hasmultipleorientationsandtheycontributealottostabilitySIFTdescriptor4.LocalimagedescriptorThresholdedimagegradientsaresampledover16x16arrayoflocationsinscalespaceCreatearrayoforientationhistograms(w.r.t.keyorientation)8orientationsx4x4histogramarray=128dimensionsNormalized,clipvalueslargerthan0.2,renormalizeσ=0.5*widthWhy4x4x8?SensitivitytoaffinechangeFeaturematchingforafeaturex,hefoundtheclosestfeaturex1
andthesecondclosestfeaturex2.Ifthedistanceratioofd(x,x1)andd(x,x1)issmallerthan0.8,thenitisacceptedasamatch.SIFTflowMaximainDRemovelowcontrastRemoveedgesSIFTdescriptorEstimatedrotationComputedaffinetransformationfromrotatedimagetooriginalimage:0.7060-0.7052128.42300.70570.7100-128.9491001.0000Actualtransformationfromrotatedimagetooriginalimage:0.7071-0.7071128.69340.70710.7071-128.6934001.0000SIFTextensionsPCAPCA-SIFTOnlychangestep4Pre-computeaneigen-spaceforlocalgradientpatchesofsize41x412x39x39=3042elementsOnlykeep20componentsAmorecompactdescriptorGLOH(Gradientlocation-orientationhistogram)17locationbins16orientationbinsAnalyzethe17x16=272-deigen-space,keep128componentsSIFTisstillconsideredthebest.SIFTMulti-ScaleOrientedPatchesSimplerthanSIFT.Designedforimagematching.[Brown,Szeliski,Winder,CVPR’2005]FeaturedetectorMulti-scaleHarriscornersOrientationfromblurredgradientGeometricallyinvarianttorotationFeaturedescriptorBias/gainnormalizedsamplingoflocalpatch(8x8)PhotometricallyinvarianttoaffinechangesinintensityMulti-ScaleHarriscornerdetectorImagestitchingismostlyconcernedwithmatchingimagesthathavethesamescale,sosub-octavepyramidmightnotbenecessary.Multi-ScaleHarriscornerdetectorsmootherversionofgradients
Cornerdetectionfunction:Picklocalmaximaof3x3andlargerthan10KeypointdetectionfunctionExperimentsshowroughlythesameperformance.Non-maximalsuppressionRestrictthemaximalnumberofinterestpoints,butalsowantthemspatiallywelldistributedOnlyretainmaximumsinaneighborhoodofradiusr.Sortthembystrength,decreasingrfrominfinityuntilthenumberofkeypoints(500)issatisfied.Non-maximalsuppressionSub-pixelrefinementOrientationassignmentOrientation=blurredgradientDescriptorVectorRotationInvariantFrameScale-spaceposition(x,y,s)+orientation(
)MSOPdescriptorvector8x8orientedpatchsampledat5xscale.SeeTRfordetails.Sampledfromwithspacing=58pixels40pixelsMSOPdescriptorvector8x8orientedpatchsampledat5xscale.SeeTRfordetails.Bias/gainnormalisation:I’=(I–
)/
Wavelettransform8pixels40pixelsDetectionsatmultiplescalesSummaryMulti-scaleHarriscornerdetectorSub-pixelrefinementOrientationassignmentbygradientsBlurredintensitypatchasdescriptorFeaturematchingExhaustivesearchforeachfeatureinoneimage,lookatalltheotherfeaturesintheotherimage(s)Hashingcomputeashortdescriptorfromeachfeaturevector,orhashlongerdescriptors(randomly)Nearestneighbortechniquesk-treesandtheirvariants(BestBinFirst)Wavelet-basedhashingComputeashort(3-vector)descriptorfroman8x8patchusingaHaar“wavelet”Quantizeeachvalueinto10(overlapping)bins(103totalentries)[Brown,Szeliski,Winder,CVPR’2005]Nearestneighbortechniquesk-Dtree
andBestBin
First
(BBF)IndexingWithoutInvariantsin3DObjectRecognition,BeisandLowe,PAMI’99ApplicationsReco
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 圣诞节创业公司活动方案
- 增加活动人气活动方案
- 外卖店经营活动方案
- 圣诞节活动创意活动方案
- 2025-2030中国数字孪生技术市场工业应用分析及标准化建设与跨行业复制潜力
- 夏天建材联盟活动方案
- 2025-2030中国人工智能芯片算力竞赛与边缘计算设备功耗优化报告
- 多元融合文旅活动方案
- 天津往年平安夜活动方案
- 大学生逻辑思维活动方案
- 公司员工公积金管理制度
- 门窗店员工管理制度
- 护士职业精神课件
- 2020年沈阳职业院校技能大赛中职学生组职业英语(服务类)样题
- 生物学基本知识
- 农业科技产业园发展战略规划与实施路径
- 2025年养老护理员(中级)考试试卷:实操技能解析
- 体育服务综合体建设项目可行性分析 (一)
- 广东深圳2025年公开招聘农村党务(村务)工作者笔试题带答案分析
- 2025-2030中国灭草松原药行业市场现状分析及竞争格局与投资发展研究报告
- 农村自建房业主培训课件
评论
0/150
提交评论