,Objectdetection,Reporter:YunshuChen2016/12/31,R-CNNvs.SPPnetvs.FastR-CNNvs.FasterR-CNN,Outline,RCNNSPPnetFastRCNNFasterRCNNExperiments,RCNN,Generatingcategoryindependentregionproposals,ExtractsafixedlengthfeaturesvectorsfromCNN,ClassspecificlinearSVMs,RegionProposals,Manydifferentalgorithmsexit.RCNNusessuperpixelbasedselectivesearch.,RCNN,ProposedasimpleandscalabledetectionalgorithmthatcombinestwoinsightsCNNisapowerfulclassifierSupervisedpretrainingfordetection,R-CNN,Image,Regions,Resize,ConvolutionFeatures,Classify,SPP-net,CNNrequiresafixedinputimagesize!,Why?,SPP-net,CNNrequiresafixedinputimagesize!,Howtodo?,SPP-net,CNNrequiresafixedinputimagesize!,SPP-netcanhandlethisissuethatRCNNSamplesanumberofboundingboxeswithdifferentsizes.,a*afeaturemap,n*nbins.TheoutputsofSPPareKM-dvectors(MisthenumberofbinsandKisthenumberofconv5filters).Thefixed-dimensionalvectorsaretheinputtothefclayer(fc6).,BoundingBox(sampledfrominputimage),CNN,Thisisimportant!,Usuallyover2000,CNNmustrunmorethan2000times,InSPPnet,convolutionrunsonce,59s/imageoncpu,SPPnet,Image,ConvolutionFeatures,SPP,Regions,Classify,R-CNNvs.SPPnet,R-CNN,SPPnet,RCNN,SPP-net,FastRCNN,RunCNNonce,Singlestagetraining,Tooslow,Multistagetraining/NoCNNfine-tuning,Takeaninputandasetofobjectproposals,FastRCNN,Generateaconvfeaturemap,ForeachBB,getafixed-lengthfeaturevectorfromROIpoolinglayerandfcs,Outputstwoinformation1)k+1classlabels2)boundingboxlocations,FastRCNN,FastRCNN,Image,ConvolutionFeatures,Regions,RoIPoolingLayer,ClassLabel,Confidence,RoIPoolingLayer,ClassLabel,Confidence,FastRCNN,R-CNNvs.SPPnetvs.FastR-CNN,R-CNN,SPPnet,FastR-CNN,RCNN,SPP-net,FastRCNN,Tooslowformultipleconvolution,Objectproposalsamplingisstillslow,Multistagetraining+NoCNNfine-tuning,FasterRCNN,SampleBBswithCNN!,RegionProposalNet+FastRCNN,FasterRCNN,RegionProposalNetwork,Inputanimageofanysize,TheregionproposalnetworkisaFCNwhichoutputsK*(4+2)sizedvectors.,Generateconvfeaturemap,Maptoalower-dimensionalfeature,Outputobjectnessscoreandboundingbox,=1,+1,LOSSPositive:AmongKanchors,onewithhighestIOU(IOU=0.7)Negative:IOU=0.3Others:Donotcontribute,RegionProposalNetwork,BBregwillhandlebadcases,Themini-batchsize(256),Thenumberofanchorlocations(2400),Image,RegionProposalNetwork,BoundingBoxRegression,BBClassification,FastR-CNN,FasterRCNN,SharingFeaturesforRPNandFastR-CNN,RPNisinitializedwithanImageNet-pre-trainedmodelandfine-tunedend-to-endfortheregionproposaltask.WetrainaseparatedetectionnetworkbyFastR-CNNusingtheproposalsgeneratedbythestep-1RPN.WeusethedetectornetworktoinitializeRPNtraining,butwefixthesharedconvolutionallayersandonlyfine-tunethelayersuniquetoRPN.Finally,keepingthesharedconvolutionallayersfixed,wefine-tunetheuniquelayersofFastR-CNN.,R-CNNvs.SPPnetvs.FastR-CNNvs.FasterR-CNN,R-CNN,SPPnet,FastR-CNN,FasterR-CNN,Experiments,Referemce,GirshickR,DonahueJ,DarrellT,etal.RichFeatureHierarchiesforAccurateObjectDetectionandSemanticSegmentationC/ComputerVisionandPatternRecognition.IEEE,2013:580-587.Cited1474HeK,ZhangX,RenS,etal.SpatialPyramidPoolinginDeepConvolutionalNetworksforVisualRecognition.J.IEEETransactionsonPatternAnalysis&MachineIntelligence,2015,37(9):1904-16.Cited312.GirshickR.FastR-CNNJ.ComputerScience,2015.Cited76.RenS,HeK,GirshickR,etal.FasterR-CNN:TowardsReal-TimeObjectDetectionwithRegionProposalNetworks.J.IEEETransactionsonPatternAnalysis&MachineIntelligence,2016:1-1.Cited184.,Thankyouforlistening!,Reporter:YunshuChen2016/12/31,