文档介绍
IEEETRANSACTIONSONPATTERNANALYSISANDMACHINEINTELLIGENCE,VOL.34,NO.2,FEBRUARY2012225AReal-TimeDeformableDetectorKarimAli,Franc¸oisFleuret,DavidHasler,andPascalFua,SeniorMember,IEEEAbstract—Weproposeanewlearningstrategyforobjectdetection.Theproposedschemeforgoestheneedtotrainacollectionofdetectorsdedicatedtohomogeneousfamiliesofposes,andinsteadlearnsasingleclassifierthathastheinherentabilitytodeformbasedonthesignalofinterest.WetrainadetectorwithastandardAdaBoostprocedurebyusingcombinationsofpose-indexedfeaturesandposeestimators.Thisallowsthelearningprocesstoselectandcombinevariousestimatesoftheposewithfeaturesabletocompensateforvariationsinposewithouttheneedtolabeldatafortrainingorexploretheposespaceintesting.Wevalidateourframeworkonthreetypesofdata:handvideosequences,aerialimagesofcars,andf