faster-rcnn

ObjectdetectionReporter:YunshuChen2016/12/31R-CNNvs.SPPnetvs.FastR-CNNvs.FasterR-CNNOutline•RCNN•SPPnet•FastRCNN•FasterRCNN•ExperimentsSelectiveSearchCNNClassifiers(SVM)Bounding-boxregressioninputproposalsfeaturesRCNN•Generatingcategoryindependentregionproposals•ExtractsafixedlengthfeaturesvectorsfromCNN•ClassspecificlinearSVMsRegionProposalsManydifferentalgorithmsexit.RCNNusessuperpixelbasedselectivesearch.SelectiveSearchCNNClassifiers(SVM)Bounding-boxregressioninputproposalsfeaturesRCNNProposedasimpleandscalabledetectionalgorithmthatcombinestwoinsights•CNNisapowerfulclassifier•SupervisedpretrainingfordetectionR-CNNImageRegionsResizeConvolutionFeaturesClassifyConvLayersSPPFcLayersClassifiers(SVM)inputFeaturemapsBounding-boxregressionSPP-netCNNrequiresafixedinputimagesize!Why?ConvLayersSPPFcLayersClassifiers(SVM)inputFeaturemapsBounding-boxregressionSPP-netCNNrequiresafixedinputimagesize!Howtodo?ConvLayersSPPFcLayersClassifiers(SVM)inputFeaturemapsBounding-boxregressionSPP-netCNNrequiresafixedinputimagesize!SPP-netcanhandlethisissuethatRCNNSamplesanumberofboundingboxeswithdifferentsizes.•a*afeaturemap,n*nbins.•TheoutputsofSPPareKM-dvectors(MisthenumberofbinsandKisthenumberofconv5filters).•Thefixed-dimensionalvectorsaretheinputtothefclayer(fc6).CNNThisisimportant!Usuallyover2000CNNmustrunmorethan2000timesInSPPnet,convolutionrunsonce59s/imageoncpuSPPnetImageConvolutionFeaturesSPPRegionsClassifyR-CNNvs.SPPnetR-CNNSPPnetRCNNSPP-netFastRCNNRunCNNonceSinglestagetrainingTooslowMultistagetraining/NoCNNfine-tuning•Betterperformance•Singlestagetraining•Allnetsaretraining•TakeaninputandasetofobjectproposalsFastRCNN•Generateaconvfeaturemap•ForeachBB,getafixed-lengthfeaturevectorfromROIpoolinglayerandfcs•Outputstwoinformation1)k+1classlabels2)boundingboxlocationsAfixed-sizevectorAfixed-sizefeaturemapROIPoolinglayersFcssoftmaxAconvFeaturemapBounding-boxregressionFcsFcFcOutput(multi-taskloss)AfullyconvolutionalnetworkObjectproposal(ROI)inputBoundingbox(similartoSPPnet)BoundingboxregressionFastRCNNFiexedsizefeaturevectorAfixed-sizevectorAfixed-sizefeaturemapROIPoolinglayersFcssoftmaxAconvFeaturemapBounding-boxregressionFcsFcFcOutput(multi-taskloss)AfullyconvolutionalnetworkObjectproposal(ROI)inputFastRCNN𝐿𝑂𝑆𝑆=𝐿𝑐𝑙𝑠𝑝,𝑢+λ[𝑢≥1]𝐿𝑙𝑜𝑐𝑡𝑢,𝑣predtragetNotbackgroundPredlocationtruelocationImageConvolutionFeaturesRegionsRoIPoolingLayerClassLabelConfidenceRoIPoolingLayerClassLabelConfidenceFastRCNNR-CNNvs.SPPnetvs.FastR-CNNR-CNNSPPnetFastR-CNNRCNNSPP-netFastRCNNTooslowformultipleconvolutionObjectproposalsamplingisstillslowMultistagetraining+NoCNNfine-tuningFasterRCNNSampleBBswithCNN!ThispartisNew!TherearesimpleFastRCNNRegionProposalNet+FastRCNNFasterRCNNRegionProposalNetwork•InputanimageofanysizeTheregionproposalnetworkisaFCNwhichoutputsK*(4+2)sizedvectors.•Generateconvfeaturemap•Maptoalower-dimensionalfeature•Outputobjectnessscoreandboundingbox𝐿𝑝𝑖,𝑡𝑖=1𝑁𝑐𝑙𝑠𝐿𝑐𝑙𝑠𝑝𝑖,𝑝𝑖∗+λ1𝑁𝑟𝑒𝑔𝑝𝑖∗𝐿𝑟𝑒𝑔𝑡𝑖,𝑡𝑖∗𝑖𝑖LOSS•Positive:AmongKanchors,onewithhighestIOU(IOU=0.7)•Negative:IOU=0.3•Others:DonotcontributeRegionProposalNetworkBBregwillhandlebadcasesThemini-batchsize(256)Thenumberofanchorlocations(~2400)ImageRegionProposalNetworkBoundingBoxRegressionBBClassificationFastR-CNNFasterRCNNSharingFeaturesforRPNandFastR-CNN•RPNisinitializedwithanImageNet-pre-trainedmodelandfine-tunedend-to-endfortheregionproposaltask.•WetrainaseparatedetectionnetworkbyFastR-CNNusingtheproposalsgeneratedbythestep-1RPN.•WeusethedetectornetworktoinitializeRPNtraining,butwefixthesharedconvolutionallayersandonlyfine-tunethelayersuniquetoRPN.•Finally,keepingthesharedconvolutionallayersfixed,wefine-tunetheuniquelayersofFastR-CNN.R-CNNvs.SPPnetvs.FastR-CNNvs.FasterR-CNNR-CNNSPPnetFastR-CNNFasterR-CNNRegionproposal(SS)Featureextraction(DeepNet)Classification(SVM)Boundingbox(Regression)RCNNRegionproposal(SS)FeatureextractionClassification+BBRegression(DeepNet)FastRCNNRegionproposalFeatureextractionClassification+BBRegression(DeepNet)FasterRCNNExperimentsReferemce•GirshickR,DonahueJ,DarrellT,etal.RichFeatureHierarchiesforAccurateObjectDetectionandSemanticSegmentation[C]//ComputerVisionandPatternRecognition.IEEE,2013:580-587.Cited1474•HeK,ZhangX,RenS,etal.SpatialPyramidPoolinginDeepConvolutionalNetworksforVisualRecognition.[J].IEEETransactionsonPatternAnalysis&MachineIntelligence,2015,37(9):1904-16.Cited312.•GirshickR.FastR-CNN[J].ComputerScience,2015.Cited76.•RenS,HeK,GirshickR,etal.FasterR-CNN:TowardsReal-TimeObjectDetectionwithRegionProposalNetworks.[J].IEEETransactionsonPatternAnalysis&MachineIntelligence,2016:1-1.Cited184.Thankyouforlistening!Reporter:YunshuChen2016/12/31