版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
BeyondBeyondGridHanHu(胡瀚)MicrosoftResearchAsiaJointworkwithJifengDai,YichenWei,ZhengZhang,JiayuanGu*,HaozhiQi*,YuwenYiLi*andGuodongBeyondGridBeyondGrid•••RelationNetworksforObjectDetection(CVPR2018oral)LearningRegionFeaturesforObjectDetection(ECCV2018)ImageImageObjectpersonInstanceInstancePose,ImageImageObjectpersonInstanceInstancePose,PersonID,CNNforImagespecificfortasksorsharedfortasksorImageCNNforImagespecificfortasksorsharedfortasksorImageGoogleNet,ResNetObjectPool+CNNforObjectFundamentalinCNNforObjectFundamentalinGoogleNet,ResNetObjectConvolutionConvolutionOutputFeature×𝐻×𝑊𝑐′,𝑐,∆ℎ,∆𝑤∙=InputFeature(𝐶𝑖𝑛×𝐻×ConvolutionConvolutionlayerisgrid/coordinateOutputConvolutionConvolutionlayerisgrid/coordinateOutputFeature×𝐻×𝑊𝑐′,𝑐,∆ℎ,∆𝑤∙=InputFeature(𝐶𝑖𝑛×𝐻×RoIPooling+FC(conv)••Itmodelspart-partRoIPooling+FC(conv)••Itmodelspart-partItiscoordinateOutputRegion(𝑁×Region(𝑁×𝐶𝑖𝑛×3×InputImage(𝐶𝑖𝑛×𝐻×FCTwoTwoImportantIssuesinComputerWithinCategoryVariation:AppearanceWithinCategoryVariation:AppearanceViewpointIntra-class(ExamplesaretakenfromLiFei-fei’scourseModelingofModelingofObject-ObjectItismucheasiertodetecttheifweknowthereisaCNNsCNNsPartlySolvetheProblem:CategoryItshouldhavenoGeometricMemorizeobjectappearanceundervariousgeometricCNNsCNNsPartlySolvetheProblem:Object-LargereceptivefieldtoincludecontextBetterBetterBeyondGridKeyIdea:Grid/CoordinateAlignment->Non-Grid/Semanticobject-objectBeyondGridKeyIdea:Grid/CoordinateAlignment->Non-Grid/Semanticobject-objectpart-partfeaturetoregionfeatureDCNfirstlearndeformationinconv&rand3rdinCOCORelationNetworksfirstmoduleforobject-objectDeformableDeformableConvolutionalJifengDai*,HaozhiQi*,YuwenXiong*,YiLi*,GuodongZhang*,HanHuandYichenWei(*EqualICCV’2017CoordinateAlignmenttoSemanticdeformableregularOutputCoordinateAlignmenttoSemanticdeformableregularOutputOutputInputInputIllustrationofDeformableOffsetsaredeterminedbyinputIllustrationofDeformableOffsetsaredeterminedbyinputdeformableBilinearEnablesbackpropagationgothroughbothoffsetbranchandinputfeaturemapsBilinearEnablesbackpropagationgothroughbothoffsetbranchandinputfeaturemapsdeformableSamplelocationandbilinearBackpropagationDeformableConvolutionisSemi-DeformableConvolutionisSemi-deformableOutputInputEfficientregularconvforwardinputcolumnEfficientregularconvforwardinputcolumnbackwardinputcolumnweightMatrixMatrixEfficientconvforwardinputcolumnEfficientconvforwardinputcolumnbackwardinputcolumnweightoffsetMatrixMatrixInput/outputInput/outputfeaturemapsarebothof…layerlayerlayerDeformable(BetterSemantic+(BetterCoordinateRoIgulardeformableRoIoutputfeatureinputfeatureInitialstateofsamplingconcurrent1KaimingHeetal.MaskR-CNN.xDeformable(BetterSemantic+(BetterCoordinateRoIgulardeformableRoIoutputfeatureinputfeatureInitialstateofsamplingconcurrent1KaimingHeetal.MaskR-CNN.xDeformableDeformableSameinput&outputastheplainRegularconvolution->deformableRegularRoIpooling->deformableRoIEnd-to-endtrainablewithoutadditionalObjectR-RoIRoIPSRoIPositionScoreR-RoIRoIPSRoIPositionScore:DeformableConvolution/RoILearnedLearnedSamplingSamplingLocationsSamplingLocationsofDeformableSamplingSamplingLocationsofDeformablePartPartOffsetsinDeformableRoIDeformableConvolutiononDeformableConvolutiononVOC&(DeepLabisaSOTAbaselineforsemanticsegmentation,whileclass-awareRPN,FasterR-CNNandR-FCNareSOTAbaselinesforobjectDeformableConvNetsonCOCO(Test-+3.3+3.1+4.7FASTERR-CNN,2FC(RESNET-CLASS-AWARERPN(RESNET-mAPDeformableConvNetsonCOCO(Test-+3.3+3.1+4.7FASTERR-CNN,2FC(RESNET-CLASS-AWARERPN(RESNET-mAP ModelComplexityModelComplexityandRuntimeonVOC&简单、粗暴(网格对齐)简单、温柔(语义对齐)、有LearningRegionLearningRegionFeaturesforObjectJiayuanGu1*,HanHu2,LiweiWang1,YichenWei2andJifeng1Peking2MicrosoftResearchAsiaRegionFeatureRegionFeatureObjectRoIRegionRoIRegionOtherOtherSpatialPyramidEnablesverylight-weighthead(per-regionBetterCoordinateAlignment+BetterSemantic…UnifiedUnifiedAGeneralExistingregionfeatureextractionAGeneralExistingregionfeatureextractionmethodsusedifferentdesignsof𝜔2(𝑏,Other…Other…Learning𝜔Learning𝜔withoutanyHandcraftedBetterencodingcontextFullyLearnableGeometric𝑊𝑏𝑜𝑥∙ ,𝑊𝑖𝑚FullyLearnableGeometric𝑊𝑏𝑜𝑥∙ ,𝑊𝑖𝑚∙<>Appearance𝑊𝑎𝑝𝑝∙Final𝜔∝exp(𝐺𝑒𝑜+SamplingforSamplingforHigherSamplingoutsideRoIregionhasalmostnoaccuracydropbuthigherExperimentsExperimentsonCOCOWhatisWhatis从从图像特征过度到区域特征的统一数学RelationRelationNetworksforObjectHanHu1*,JiayuanGu2*,ZhengZhang1*,JifengDai1,andYichen1MicrosoftResearchAsia2PekingUniversityRelationModelinginpart-partEffectiveRelationModelinginpart-partEffectiveandEasytoRequirenorelationTranslationalObject-objectrelation:wellrecognizedpart-partobject-objectObject-objectrelation:wellrecognizedpart-partobject-object?WellRecognizedWellRecognizedItismucheasiertodetecttheifweknowthereisaRarelyStudiedinDeepLearningRarelyStudiedinDeepLearningIrregularitiesofAtarbitraryimageOfdifferentWithindifferentOfvaryingnumberacrossdifferentimagesGoal:designGoal:designasimplemoduletoobject-objectEffectiveandEasytoRequirenorelationIn-place,ObjectRelationObjectRelationExtensionofObjectRelationExtensionofattentionobject-object(2D(1DLeftfigurecreditbyA.VaswanietRelationbetweenTwoAnovelgeometricappearanceRelationbetweenTwoAnovelgeometricappearanceapp.+geometricdotsmallmax{0,𝑊𝐺∙(𝑞)4dboundingregressioninstandardattentioninobjectrelationRelation𝑜𝑢𝑡(𝑛)appearance+Relation𝑜𝑢𝑡(𝑛)appearance+geometric𝑓𝑜𝑢𝑡(𝑛)=ω(𝑚,𝑛)𝑛𝑚)objectMulti-Branchbranch(person-branch(playground-…Multi-Branchbranch(person-branch(playground-…branch(duplicate…ObjectRelationresidualoutput=ObjectRelationresidualoutput=input+…KaimingHe,XiangyuZhang,ShaoqingRen,JianSun.DeepResidualLearningforImageRecognition.CVPR,ObjectRelationEffectiveandObjectRelationEffectiveandEasytoRequirenorelationTranslational…Application:Application:ObjectRegion-basedObjectR-regionRegion-basedObjectR-regionR.Girshick.FastR-CNN.ICCV,S.Renetal.FasterR-CNN.NIPS,Ourmethod:insertingobjectrelationmodulesindependenthandcaftOurmethod:insertingobjectrelationmodulesindependenthandcaftexR-learnable××InstanceDuplicateLearnableDuplicate√LearnableDuplicate√originalTheFirstFullyEnd-to-EndObject×TheFirstFullyEnd-to-EndObject××backpropagation××*Faster××*FasterR-CNNwithResNet-50modelare+2.3mAPbyinserting2with+3%××××*FasterR-CNNwithResNet-50modelareMoremodules:8××*Faster××*FasterR-CNNwithResNet-50modelareImportanceofrelativegeometric××*Faster××*FasterR-CNNwithResNet-50modelareImportanceofmulti-branch××*Faster××*FasterR-CNNwithResNet-50modelareImportanceofresidualDuplicateRemovalNoticeablybetterthanDuplicateRemovalNoticeablybetterthanSlightlybetterthanSoftNMS[N.Bodlaetal,2017]TrainableFullyEnd-to-EndObjectBenefitfromFullyEnd-to-EndObjectBenefitfromfullyend-to-endTrainableUsingStronger+3.0UsingStronger+3.0+2.0+1.0*FasterR-CNNwithResNet-101modelareused(
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 2024年山东客运从业资格证实际操作试题及答案
- 2024年信用借款合同签完多久放款(2730字)
- 2024年标准钢材购销合同书范本
- 港航实务 皮丹丹 教材精讲班课件 79-第6章-6.1-工程招标投标
- 2024年五金电器购销简单合同范本
- 2024年品质保证协议书转让协议
- 2024年水电承包合同
- 2024年经营合作协议范本
- 2024年简易个人房屋装修合同
- 2024年杂志约稿合同模板
- 新视野大学英语视听说教程ppt课件
- 攻城掠地数据以及sdata文件修改教程
- 医疗废物转运箱消毒记录表
- 最新投标书密封条
- 看守所岗位职责
- 2019年青年英才培养计划项目申报表
- Sentaurus在ESD防护器件设计中的应用PPT课件
- 《抛物线焦点弦的性质探究》学案
- 人教版小学二年级数学上册全册教案【表格式】
- 佛山岭南新天地项目概况.
- 喷码机操作手册
评论
0/150
提交评论