Deformable Part Models are Convolutional Neural Networks

《Deformable Part Models are Convolutional Neural Networks》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、DeformablePartModelsareConvolutionalNeuralNetworksRossGirshick1ForrestIandola2TrevorDarrell2JitendraMalik21MicrosoftResearch2UCBerkeleyrbg@microsoft.comfforresti,trevor,malikg@eecs.berkeley.eduAbstractCNN.Inotherwords,deformablepartmodelsareconvo-lutionalneuralnetworks.Ourconstructionrelieso

2、nanewDeformablepartmodels(DPMs)andconvolutionalneu-networklayer,distancetransformpooling,whichgeneral-ralnetworks(CNNs)aretwowidelyusedtoolsforvi-izesmaxpooling.sualrecognition.Theyaretypicallyviewedasdistinctap-DPMstypicallyoperateonascale-spacepyramidofgra-proaches:DPMsaregraphicalmodels(M

3、arkovrandomdientorientationfeaturemaps(HOG[5]).Butwenowfields),whileCNNsare“black-box”non-linearclassifiers.knowthatforobjectdetectionthisfeaturerepresentationisInthispaper,weshowthataDPMcanbeformulatedasasuboptimalcomparedtofeaturescomputedbydeepcon-CNN,thusprovidingasynthesisofthetwoideas.Ou

4、rcon-volutionalnetworks[17].Asasecondinnovation,were-structioninvolvesunrollingtheDPMinferencealgorithmplaceHOGwithfeatureslearnedbyafully-convolutionalandmappingeachsteptoanequivalentCNNlayer.Fromnetwork.This“front-end”networkgeneratesapyramidofthisperspective,itisnaturaltoreplacethestandar

5、dim-deepfeatures,analogoustoaHOGfeaturepyramid.WeagefeaturesusedinDPMswithalearnedfeatureextractor.callthefullmodelaDeepPyramidDPM.WecalltheresultingmodelaDeepPyramidDPMandex-WeexperimentallyvalidateDeepPyramidDPMsbyperimentallyvalidateitonPASCALVOCobjectdetection.measuringobjectdetectionper

6、formanceonPASCALVOCWefindthatDeepPyramidDPMssignificantlyoutperform[9].SincetraditionalDPMshavebeentunedforHOGfea-DPMsbasedonhistogramsoforientedgradientsfeaturesturesovermanyyears,wefirstanalyzethedifferencesbe-(HOG)andslightlyoutperformsacomparableversionoftweenHOGfeaturepyramidsanddeepfeatur

7、epyramids.therecentlyintroducedR-CNNdetectionsystem,whilerun-WethenselectagoodmodelstructureandtrainaDeep-ningsignificantlyfaster.PyramidDPMthatsignificantlyoutperformsthebestHOG-basedDPMs.Whilewedon’texpectourapproachtoout-performafine-tunedR-CNNdete