利用空间上下文做物件识别

利用空间上下文做物件识别

ID:5405721

大小:11.88 MB

页数:42页

时间:2017-11-10

利用空间上下文做物件识别_第1页
利用空间上下文做物件识别_第2页
利用空间上下文做物件识别_第3页
利用空间上下文做物件识别_第4页
利用空间上下文做物件识别_第5页
资源描述:

《利用空间上下文做物件识别》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、利用空间上下文寻找识别物体原作者:GeremyHeitz,DaphneKollerStanfordUniversityThingsvs.StuffStuff(n):Materialdefinedbyahomogeneousorrepetitivepatternoffine-scaleproperties,buthasnospecificordistinctivespatialextentorshape.Thing(n):Anobjectwithaspecificsizeandshape.From:Forsyth

2、etal.Findingpicturesofobjectsinlargecollectionsofimages.ObjectRepresentationinComputerVision,1996.FindingThingsContextiskey!OutlineWhatisContext?TheThingsandStuff(TAS)modelResultsSatelliteDetectionExampleD(W)=0.8D(W)=0.8ErrorAnalysisTypically…Weneedtolookout

3、sidetheboundingbox!FalsePositivesare OUTOFCONTEXTTruePositivesare INCONTEXTTypesofContextScene-Thing:Stuff-Stuff:gistcar“likely”keyboard“unlikely”Thing-Thing:[Torralbaetal.,LNCS2005][Gouldetal., IJCV2008][Rabinovichetal.,ICCV2007]TypesofContextStuff-Thing:Ba

4、sedonspatialrelationshipsIntuition:Trees=nocarsHouses=carsnearbyRoad=carshere“Carsdriveonroads”“Cowsgrazeongrass”“Boatssailonwater”Goal:UnsupervisedOutlineWhatisContext?TheThingsandStuff(TAS)modelResultsThingsDetection“candidates”Lowdetectorthreshold->“over-

5、detect”EachcandidatehasadetectorscoreThingsCandidatedetectionsImageWindowWi+ScoreBooleanR.V.TiTi=1:CandidateisapositivedetectionThingmodelTiImage WindowWiStuffCoherentimageregionsCoarse“superpixels”FeaturevectorFjinRnClusterlabelSjin{1…C}StuffmodelNaïveBayes

6、SjFjRelationshipsDescriptiveRelations“Near”,“Above”, “Infrontof”,etc.ChoosesetR={r1…rK}Rijk=1:DetectioniandregionjhaverelationkRelationshipmodelS72=TreesS4=HousesS10=RoadT1RijkTiSjR1,10,in=1TheTASModelRijkTiSjFjImage WindowWiWi:WindowTi:ObjectPresenceSj:Regi

7、onLabelFj:RegionFeaturesRijk:RelationshipNJKSupervised inTrainingSetAlways ObservedAlways HiddenUnrolledModelT1S1S2S3S4S5T2T3R2,1,above=0R3,1,left=1R1,3,near=0R3,3,in=1R1,1,left=1Candidate WindowsImage RegionsLearningtheParametersAssumeweknowRSjishiddenEvery

8、thingelseobservedExpectation-Maximization“Contextualclustering”ParametersarereadilyinterpretableRijkTiSjFjImage WindowWiNJKSupervised inTrainingSetAlways ObservedAlways HiddenLearnedSatelliteClu

当前文档最多预览五页,下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页,下载文档查看全文
温馨提示:
1. 部分包含数学公式或PPT动画的文件,查看预览时可能会显示错乱或异常,文件下载后无此问题,请放心下载。
2. 本文档由用户上传,版权归属用户,天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容,确认文档内容符合您的需求后进行下载,若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误,付费完成后未能成功下载的用户请联系客服处理。