wenshu-weely report-Shape Matching and Object Reco

ShapeMatchingandObjectRecognitionusingShapeContextsweeklyreportWenshuLi2010.07.03FrameworkonobjectdetectionOutline•Objectrecognition(Shapematchingandisolatedobjectrecognition)•ScalinguptogeneralobjectrecognitionBiologicalShape•D’ArcyThompson:OnGrowthandForm,1917–studiedtransformationsbetweenshapesoforganismsHumanoranimalskeletonMatchingFramework•Findcorrespondencesbetweenpointsonshape•Estimatetransformation•Measuresimilaritymodeltarget...ComparingPointsetsShapeContextCountthenumberofpointsinsideeachbin,e.g.:Count=4Count=10...CompactrepresentationofdistributionofpointsrelativetoeachpointShapeContextShapeContexts•Invariantundertranslationandscale•Canbemadeinvarianttorotationbyusinglocaltangentorientationframe•Toleranttosmallaffinedistortion–Log-polarbinsmakespatialblurproportionaltorCf.SpinImages(Johnson&Hebert)-rangeimageregistrationComparingShapeContextsComputematchingcostsusingChiSquareddistance:RecovercorrespondencesbysolvinglinearassignmentproblemwithcostsCij[Jonker&Volgenant1987]MatchingFramework•Findcorrespondencesbetweenpointsonshape•Estimatetransformation•Measuresimilaritymodeltarget...MatchingFramework•Findcorrespondencesbetweenpointsonshape•Estimatetransformation•Measuresimilaritymodeltarget...Outline•Shapematchingandisolatedobjectrecognition•Scalinguptogeneralobjectrecognition–Manyobjects–Grayscalematching–Objectsinscenes(scanningorsegmentation)Manyobjects•FastPruning–Givenaqueryshape,quicklyreturnashortlistofcandidatematches–Databaseofknownobjectswillbelarge:~30000•DetailedMatching–PerformcomputationallyexpensivecomparisonsononlythefewshapesintheshortlistRepresentativeShapeContexts•Matchusingonlyafewshapecontexts–Don’tneedtocompareeveryone2*12*(,)(,)argmin(,)rjqueryiqueryijjuiuqueryidistSSSCSCSCSCSCResults•codebookdataTheCodebookEntries(CE)arearepositoryofexamplefeatures:CE={cei}.Eachcodebookentrycei={ui,deltai,mi,wi}recordsthefeatureforapointiinlabeledobjectsofthetrainingimages.Hereuiistheshapecontextvectorforpointi.deltaiisthepositionofpointirelativetotheobjectcenter.miisabinarymaskoffigure-groundsegmentationforthepatchcenteredatpointi.wiistheweightmaskcomputedonmi.DetectionacrossscalesThefirstcolumnisedgedetectionresults.Thesecondcolumnisvotingresults.Thethirdcolumnisselectedhypothesispositionandboundaryboxs.Fig1.InputimageFig2.Edgedetectionresultofinputimage.Fig3.Samplepointsoninputimagetoextractfeatureson.Fig4.'A'isapointoninputimageFig5.These7imagesarefromcodebook,whicharemodelimages.'B'-'H'arematchesto'A'inFig4;Eachmatchwillindicatewherethebodycentercouldpossiblybe.Fig6.Voteresult.HighvalueindicatehighpossibilitythatpositioncouldbeahypothesiscenterFig7.Selecthypothesespositionfromvotemap(Fig.6),indicatedbymagentacrossFig8.Foreachhypothesis,tracebacktofinditsvotersandeachvotepredicttheforeground/backgroundlabelingoftheinputimage,whichgivesahypothesismask,andboundaryboxscanbeeasilyderivedfromthismaskFig9.Afterforceone-to-oneuniquematching,wegetannewscoreforeachhypothesis56.88552.396124.455524.261621.5206501001502002503003504004505005505010015020025030035040045050028.120422.621217.281917.001410020030040050060050100150200250300350400450Futurework•Studynewmatchingalgorithmmatchingcorrespondingtoshapecontexts•Robusttooutliers&noise•Formsbasisofobjectrecognitiontechniquethatperformswellinavarietyofdomainsusingexactlythesamealgorithm