用于物体检测字典对分类器驱动卷积神网络dictionary pair classifier cvpr paperwang.pdfVIP

DictionaryPairClassifierDrivenConvolutionalNeuralNetworks

forObjectDetection

1,31∗233

KezeWang,LiangLin,WangmengZuo,ShuhangGu,LeiZhang

1SchoolofDataandComputerScience,SunYat-senUniversity,Guangzhou,

2SchoolofComputerScienceandTechnology,HarbinInstituteofTechnology,

3DepartmentofComputing,ThePolytechnicUniversity

linliang@;{kezewang,cswmzuo}@;{cssgu,cslzhang}@.hk.

tractinggreatinterestsincomputervisioncommunity.Al-

thoughitsperformancehasbeenimprovedsubstantiallyin

Featurerepresentationandobjectcategoryclassifi-thepastdecade[31,34,21,8,30,33,13,16],objectdetec-

cationaretwokeycomponentsofmostobjectdetec-tionremainsachallengeproblemundercomplexandun-

tionmethods.Whilesignificantimprovementshavebeenconstrainedenvironments.

achievedfordeepfeaturerepresentationlearning,tradi-

Recently,groundbreakingprogressonobjectdetection

tionalSVM/softmaxclassifiersremainthedominantmeth-

hasbeenmadeduetotheadvancesindeepconvolutional

odsforthefinalobjectcategoryclassification.However,

neuralnetworks(CNNs)[19,28]andtheincreasingsize

SVM/