Maximum margin multi-instance learning_NIPS2011

《Maximum margin multi-instance learning_NIPS2011》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、MaximumMarginMulti-InstanceLearningHuaWangHengHuangComputerScienceandEngineeringComputerScienceandEngineeringUniversityofTexasatArlingtonUniversityofTexasatArlingtonhuawangcs@gmail.comheng@uta.eduFarhadKamangarFeipingNieComputerScienceandEngineeringComputerScienc

2、eandEngineeringUniversityofTexasatArlingtonUniversityofTexasatArlingtonkamangar@uta.edufeipingnie@gmail.comChrisDingComputerScienceandEngineeringUniversityofTexasatArlingtonchqding@uta.eduAbstractMulti-instancelearning(MIL)considersinputasbagsofinstances,inwhichl

3、a-belsareassignedtothebags.MILisusefulinmanyreal-worldapplications.Forexample,inimagecategorizationsemanticmeanings(labels)ofanimagemostlyarisefromitsregions(instances)insteadoftheentireimage(bag).ExistingMILmethodstypicallybuildtheirmodelsusingtheBag-to-Bag(B2B)

4、distance,whichareoftencomputationallyexpensiveandmaynottrulyreflectthesemanticsim-ilarities.Totacklethis,inthispaperweapproachMILproblemsfromanewperspectiveusingtheClass-to-Bag(C2B)distance,whichdirectlyassessestherelationshipsbetweentheclassesandthebags.Takingint

5、oaccountthetwoma-jorchallengesinMIL,highheterogeneityondataandweaklabelassociation,weproposeanovelMaximumMarginMulti-InstanceLearning(M3I)approachtoparameterizetheC2Bdistancebyintroducingtheclassspecificdistancemetricsandthelocallyadaptivesignificancecoefficients.We

6、applyournewapproachtotheautomaticimagecategorizationtasksonthree(onesingle-labelandtwomulti-label)benchmarkdatasets.Extensiveexperimentshavedemonstratedpromisingresultsthatvalidatetheproposedmethod.1IntroductionTraditionalimagecategorizationmethodsusuallyconsider

7、animageasoneindiscreteentity,which,however,neglectsanimportantfactthatthesemanticmeanings(labels)ofanimagemostlyarisefromitsconstituentregions,butnottheentireimage.Forexample,thelabels“person”and“car”associatedwiththequeryimageinFigure1areonlycharacterizedbythere

8、gionsintwoboundingboxes,respectively,ratherthanthewholeimage.Therefore,modelingtherelationshipsbetweenla-belsandregions(insteadoftheentireimage)couldpotentiall