MRFs与图割法在计算机视觉中应用及分割技术.pptx

MRFsandSegmentationwithGraphCutsComputerVisionCS543/ECE549UniversityofIllinoisDerekHoiem03/31/15

Logistics HW3gradesreleasedHW4duenextMondayFinalprojectproposalsdueFriday(byemail)

Today’sclassReview/FinishEMMRFsSegmentationwithGraphCutsijwij

MissingDataProblems:Segmentation Challenge:Segmenttheimageintofigureandgroundwithoutknowingwhattheforegroundlookslikeinadvance.Threesubproblems:Ifwehadlabels,howcouldwemodeltheappearanceofforegroundandbackground? MLE:maximumlikelihoodestimationOncewehavemodeledthefg/bgappearance,howdowecomputethelikelihoodthatapixelisforeground?ProbabilisticinferenceHowcanwegetbothlabelsandappearancemodelsatonce? Hiddendataproblem:ExpectationMaximizationForegroundBackground

EM:MixtureofGaussiansInitializeparametersComputelikelihoodofhiddenvariablesforcurrentparametersEstimatenewparametersforeachcomponent,weightedbylikelihoodGivenbynormaldistribution

GaussianMixtureModels:PracticalTipsDesigndecisionsNumberofcomponentsSelectbyhandbasedonknowledgeofproblemSelectusingcross-validationorsampledataUsually,nottoosensitiveandsafertousemorecomponentsVariance“Sphericalcovariance”:dimensionswithineachcomponentareindependentwithequalvariance(1parameterbutusuallytoorestrictive)“Diagonalcovariance”:dimensionswithineachcomponentarenotindependentwithdifferencevariances(NparametersforN-Ddata)“Fullcovariance”:noassumptions(N*(N+1)/2parameters);forhighNmightbeexpensivetodoEM,toevaluate,andmayoverfitTypicallyuse“Full”iflotsofdata,fewdimensions;Use“Diagonal”otherwiseCangetstuckinlocalminimaUsemultiplerestartsChoosesolutionwithgreatestdatalikelihood

“HardEM”SameasEMexceptcomputez*asmostlikelyvaluesforhiddenvariablesK-meansisanexampleAdvantagesSimpler:canbeappliedwhencannotderiveEMSometimesworksbetterifyouwantt