[ECCV 2010] Clustering Complex Data with Group-Dependent Feature Selection

《[ECCV 2010] Clustering Complex Data with Group-Dependent Feature Selection》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、ClusteringComplexDatawithGroup-DependentFeatureSelectionYen-YuLin1,2,Tyng-LuhLiu1,andChiou-ShannFuh21InstituteofInformationScience,AcademiaSinica,Taipei,Taiwan2DepartmentofCSIE,NationalTaiwanUniversity,Taipei,Taiwan{yylin,liutyng}@iis.sinica.edu.tw,fuh@csi

2、e.ntu.edu.twAbstract.Wedescribeaclusteringapproachwiththeemphasisonde-tectingcoherentstructuresinacomplexdataset,andillustrateitseffec-tivenesswithcomputervisionapplications.Bycomplexdata,wemeanthattheattributevariationsamongthedataaretooextensivesuchthatcl

3、usteringbasedonasinglefeaturerepresentation/descriptorisinsuffi-cienttofaithfullydividethedataintomeaningfulgroups.Theproposedmethodthusassumesthedataarerepresentedwithvariousfeaturerep-resentations,andaimstouncovertheunderlyingclusterstructure.Tothatend,wea

4、ssociateeachclusterwithaboostingclassifierderivedfrommultiplekernellearning,andapplythecluster-specificclassifiertofeatureselectionacrossvariousdescriptorstobestseparatedataoftheclusterfromtherest.Specifically,weintegratethemultiple,correlativetrainingtasksoft

5、hecluster-specificclassifiersintotheclusteringprocedure,andcastthemasajointconstrainedoptimizationproblem.Throughtheop-timizationiterations,theclusterstructureisgraduallyrevealedbytheseclassifiers,whiletheirdiscriminantpowertocapturesimilardatawouldbeprogress

6、ivelyimprovedowingtobetterdatalabeling.1IntroductionClusteringisatechniquetopartitionthedataintogroupssothatsimilar(orcoherent)objectsandtheirpropertiescanbereadilyidentifiedandexploited.Whilesuchagoalisexplicitandclear,thenotionofsimilarityisoftennotwellde

7、fined,partlyduetothelackofauniversallyapplicablesimilaritymeasure.Asaresult,previousresearcheffortsondevelopingclusteringalgorithmsmostlyfocusondealingwithdifferentscenariosorspecificapplications.Inthefieldofvisionresearch,performingdataclusteringisessentialina

8、ddressingvarioustaskssuchasobjectcategorization[1,2]orimagesegmentation[3,4].Despitethegreatapplicability,afundamentaldifficultyhinderingtheadvanceofclusteringtechniquesisthattheintrinsicclusterstructureisnotev