弱监督深度检测网络牛津hbilenrobots ox ac uk andrea weakly supervised deep cvpr paper.pdf

WeaklySupervisedDeepDetectionNetworks

HakanBilenAndreaVedaldi

UniversityofOxfordUniversityofOxford

hbilen@robots.ox.ac.ukvedaldi@robots.ox.ac.uk







Weaklysupervisedlearningofobjectdetectionisanim-

portantprobleminimageunderstandingthatstilldoesnot

haveasatisfactorysolution.Inthispaper,weaddressthis

problembyexploitingthepowerofdeepconvolutionalneu-



ralnetworkspre-trainedonlarge-scaleimage-levelclassi-



ficationtasks.Weproposeaweaklysuperviseddeepdetec-

tionarchitecturethatmodifiesonesuchnetworktooperate

atthelevelofimageregions,performingsimultaneouslyre-

gionselectionandclassification.Trainedasanimageclas-

sifier,thearchitectureimplicitlylearnsobjectdetectorsthatFigure1.WeaklySupervisedDeepDetectionNetwork.Our

arterthanalternativeweaklysuperviseddetectionsys-methodstartsfromaCNNpre-trainedf