斯坦福大学计算机视觉课件

PatternRecognitionforVisionFall2004Vision—FeatureExtractionTopicsPartI•FourierTransform•WindowedFourierTransform•WaveletsPartII•PrincipalComponentAnalysis•IndependentComponentAnalysisPatternRecognitionforVisionFall2004Vision—FeatureExtractionI2()ˆ(,)(,)xyyjxyxffxyedxdyπωωωω∞∞−+−∞−∞=∫∫FourierTransformxyImage(B)(A)SpectrumxωyωAmplitudeof(A)&Phaseof(B)PhasecarriesInformationCourtesyofProfessorsTomasoPoggioandSayanMukherjee.Usedwithpermission.PatternRecognitionforVisionFall2004Vision—FeatureExtractionITemplateMatching'''()()()()()cxfxgxxdxhxgx∞−∞=−=−∫'''ˆˆˆ()()()()()()cxfxhxxdxcfhωωω∞−∞=−=∫ResultcTemplategImagefPatternRecognitionforVisionFall2004Vision—FeatureExtractionIWindowedFourierTransform2(,)()()juftfuguteduπωω∞−−∞=−∫%2inst()cos(),ftttπω==()gt(,)ftω%()fuu()gut−tT−tPatternRecognitionforVisionFall2004Vision—FeatureExtractionIWaveletTransform,1()stputussψψ−=,(,)()()stTstfuuduψ∞−∞=∫20.5,101,02,15()redgreenblueuuueψψψψ−−−−=PatternRecognitionforVisionFall2004Vision—FeatureExtractionIHaarWavelets(MatlabToolbox)ScreenshotfromMatlabToolboxremovedduetocopyrightreasons.PatternRecognitionforVisionFall2004Vision—FeatureExtractionIIPrincipalandIndependentComponentAnalysisx1PCAICAPCAICAx2PCA:•Decorrelated•OrthogonalICA:•StatisticallyIndependentImageremovedduetocopyrightconsiderations.SeeFigure1in:Baek,Kyungim,et.al.PCAvs.ICA:AcomparisonontheFERETdataset.InternationalConferenceofComputerVision,PatternRecognition,andImageProcessing,inconjunctionwiththe6thJCIS.Durham,NC,March8-142002,June2001.PatternRecognitionforVisionFall2004PatternRecognition—ClassifiersTopicsFisherDiscriminantAnalysis(FDA)•SupportVectorMachines(SVM)PatternRecognitionforVisionFall2004PatternRecognition—ClassifiersFisherDiscriminantAnalysisx2Maximize2122212()mmσσ−+x1PatternRecognitionforVisionFall2004PatternRecognition—ClassifiersLinearSupportVectorMachines1()Niiiifsignybα==⋅−∑xxxMaximizemargindixPatternRecognitionforVisionFall2004PatternRecognition—ClassifiersNon-linearSupportVectorMachinesx1x2InputSpacex*1x*2)(*xxΦ=FeatureSpace1()(,),(,)()()NiiiiiifsignyKbKα==−=ΦΦ∑xxxxxxxPatternRecognitionforVisionFall2004Vision—BiologicalObjectRecognitionTopics•VisualCortex•HierarchicalProcessing•ScaleInvariance•MAXModelforObjectRecognitionPatternRecognitionforVisionFall2004Vision—BiologicalObjectRecognitionTheVisualSystemSimplifiedVentralstream:What?Figure1in:Ungerleider,L.,et.al.Aneuralsystemforhumanvisualworkingmemory.ProceedingsoftheNationalAcademyofSciences95(February1998):883-890.Copyright1998NationalAcademyofSciences,U.S.A.PatternRecognitionforVisionFall2004Vision—BiologicalObjectRecognitionFromSmallandSimpletoBigandComplexMax’sstuffV1V4ITCmodifiedfromUngerleiderandHaxby,1994ITcells:ObjectspecificHubel&Wiesel,1959Simplecells:OrientationofbarsImageremovedduetocopyrightconsiderations.See:Hubel,andWeisel.Receptivefieldsofsingleneuronsinthecat'sstriatecortex.JournalofPhysiology148(1959):574-951.Imageremovedduetocopyrightconsiderations.See:Ungerleider,andHaxby.'What'and'where'inthehumanbrain.CurrentOpinioninNeurobiology4,no.2(1994):157-165.PatternRecognitionforVisionFall2004Vision—BiologicalObjectRecognitionHierarchicalProcessing:FromSimpletoComplexFeaturesFigure1in:Wersing,H.,andE.Korner.LearningOptimizedFeaturesforHierarchicalModelsofInvariantObjectRecognition.NeuralComputation15,no.7(2003):1559-1558.MITPressJournals.Courtesyof________________MITPressJournals.Usedwithpermission.PatternRecognitionforVisionFall2004Applications—MorphableModelsTopics•2DMMsoffacicalcomponents•2DMMsforfacialanimation•3DMMsoffaces•ModifyingfacesintheMMspacePatternRecognitionforVisionFall2004Applications—MorphableModels2DMorphableModelsforFacialAnimation“Air”“Badge”VideoDatabaseVisualSpeechProcessing“Badge”•Videorealism•Machinelearning1I2I3I4I2C3C4CBuildinga2DMMFromFigure1in:Ezzat,Geiger,andPoggio.TrainableVideorealisticSpeechAnimation.ProceedingsofSIGGRAPH2002,SanAntonio,TX.Courtesyofauthors.Usedwithpermission.PatternRecognitionforVisionFall2004Applications—MorphableModels2DMorphableModelsforFacialAnimationFromFigure10in:Ezzat,Geiger,andPoggio.TrainableVideorealisticSpeechAnimation.ProceedingsofSIGGRAPH2002,SanAntonio,TX.Courtesyofauthors.Usedwithpermission.PatternRecognitionforVisionFall2004Applications—MorphableModelsGenerating3DFaceModelswithMMFaceImages3DFaceModelMorphingSpaceCourtesyofMaxmillianRiesen-Huber.Usedwithpermission.PatternRecognitionforVisionFall2004Applications—MorphableModelsModifyingFacesintheMMSpaceNovelImagesFiguresremovedduetocopyrightconsiderations.Pleasesee:FiguresfromVetter,T.,andV.Blanz.AMorphableModelfortheSynthesisof3DFaces.ProceedingsofSIGGRAPH(1999).PatternRecognitionforVisionFall2004Applications—ObjectDetection&RecognitionTopics•Facedetection&recognition•Pedestriandetection•Featureextraction•ClassificationPatternRecognitionforVisionFall2004Applications—ObjectDetection&RecognitionObjectDetectionSearchforfacesatdifferentresolutionsandlocationsFeaturevector(x1,x2,…,xn)ClassifierOff-linetrainingFaceexamplesNon-faceexamplesFeatureExtractionPixe