伊利诺伊州立大学计算机视觉PPT第23章.pdf

《伊利诺伊州立大学计算机视觉PPT第23章.pdf》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、04/10/12ObjectCategoryDetection:SlidingWindowsComputerVisionCS543/ECE549UniversityofIllinoisDerekHoiemToday’sclass:ObjectCategoryDetection•Overviewofobjectcategorydetection•Statisticaltemplatematchingwithslidingwindowdetector–Dalal-Triggspedestriandetector–

2、Viola-JonesfacedetectorObjectCategoryDetection•Focusonobjectsearch:“Whereisit?”•BuildtemplatesthatquicklydifferentiateobjectpatchfrombackgroundpatchDogModelObjectorNon-Object?ChallengesinmodelingtheobjectclassIlluminationObjectposeClutterOcclusionsIntra-cla

3、ssViewpointappearanceSlidefromK.Grauman,B.LeibeChallengesinmodelingthenon-objectclassTrueDetectionsBadConfusedwithLocalizationSimilarObjectConfusedwithMisc.BackgroundDissimilarObjectsGeneralProcessofObjectRecognitionWhataretheobjectSpecifyObjectModelparamet

4、ers?GenerateHypothesesScoreHypothesesResolveDetectionsSpecifyinganobjectmodel1.StatisticalTemplateinBoundingBox–Objectissome(x,y,w,h)inimage–FeaturesdefinedwrtboundingboxcoordinatesImageTemplateVisualizationImagesfromFelzenszwalbSpecifyinganobjectmodel2.Art

5、iculatedpartsmodel–Objectisconfigurationofparts–EachpartisdetectableImagesfromFelzenszwalbSpecifyinganobjectmodel3.Hybridtemplate/partsmodelDetectionsTemplateVisualizationFelzenszwalbetal.2008Specifyinganobjectmodel4.3D-ishmodel•Objectiscollectionof3Dplanar

6、patchesunderaffinetransformationGeneralProcessofObjectRecognitionSpecifyObjectModelProposeanalignmentoftheGenerateHypothesesmodeltotheimageScoreHypothesesResolveDetectionsGeneratinghypotheses1.Slidingwindow–TestpatchateachlocationandscaleGeneratinghypothese

7、s1.Slidingwindow–TestpatchateachlocationandscaleGeneratinghypotheses2.Votingfrompatches/keypointsMatchedCodebookProbabilisticInterestPointsEntriesVotingysx3DVotingSpace(continuous)ISMmodelbyLeibeetal.Generatinghypotheses3.Region-basedproposalEndresHoiem2010

8、GeneralProcessofObjectRecognitionSpecifyObjectModelGenerateHypothesesMainly-gradientbasedfeatures,usuallybasedonScoreHypothesessummaryrepresentation,manyclassifiersResolveDetectionsGeneralProcessofObje