资源描述：
《Segmentation-UniversityofMarylandInstituteforAdvanced分割-先进的马里兰大学研究所》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。
1、SegmentationKyongilYoonSegmentationObtainacompactrepresentationofwhatishelpful(intheimage)NocomprehensivetheoryofsegmentationHumanvision:GroupingandGestaltProximity,similarity,commonfate,commonregion,parallelism,closure,symmetry,continuity,familiarconfigurationSegmentationbycluster
2、ingPartitioningvs.groupingApplicationsBackgroundsubtractionShotboundarydetectionImagesegmentationbyclusteringpixelsUsingsimpleclusteringAgglomerativeclustering(clusteringbymerging)Divisiveclustering(clusteringbysplitting)K-meansUsinggraph-theoreticclusteringAffinitymeasureNormalize
3、dcutcut(A,B)/assoc(A,V)+cut(A,B)/assoc(B,V)K-MeansChoosekdatapointstoactasclustercentersUntiltheclustercentersareunchangedAllocateeachdatapointtoclusterwhosecenterisnearestNowensurethateveryclusterhasatleastonedatapoint;possibletechniquesfordoingthisincludesupplyingemptyclusterswit
4、hapointchosenatrandomfrompointsfarfromtheirclustercenter.Replacetheclustercenterswiththemeanoftheelementsintheirclusters.endGraphEigenvectorsConstructanaffinitymatrixComputetheeigenvaluesandeigenvectorsoftheaffinitymatrixUntiltherearesufficientclustersTaketheeigenvectorcorrespondin
5、gtothelargestunprocessedeigenvalue;zeroallcomponentscorrespondingtoelementsthathavealreadybeenclustered,andthresholdtheremainingcomponentstodeterminewhichelementbelongstothiscluster,choosingathresholdbyclusteringthecomponents,orusingathresholdfixedinadvance.Ifallelementshavebeenacc
6、ountedfor,therearesufficientclustersendSegmentationbyfittingamodelToassertthatpixelsbelongtogethertoconformtosomemodelLargescaleexplicitmodelHoughtransformThreeproblems:Whatistheline?Whichpointsbelongtowhichline?Howmanylines?Pointspace<->linespacex*cos(t)+y*sin(t)+r=0,(t,r)linespac
7、eHalf-infinitecylinderQuantizationerrors,difficultieswithnoiseFittinglines,fittingcurvesLeastsquareTotalleastsquareSegmentationbyfittingamodel(2)TwobigproblemsRobustness:whatifonedatapointisFAR,andallothersfillwell?Missingdata:whichpointisnoiseandwhichpointisnot?RobustnessOutliers:
8、Improvethemodeleitherbygiv
