资源描述：
《Ch04 High-Level MRF Models》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、Chapter4High-LevelMRFModelsHigh-levelvisiontasks,suchasobjectmatchingandrecognitionandposeestimation,areperformedonfeaturesextractedfromimages.Thearrange-mentsofsuchfeaturesareusuallyirregular,andhencetheproblemsfallintocategoriesLP3andLP4.Inthischapter,wepresentMAP-MRFformulationsfo
2、rsolvingtheseproblems.Webeginwithastudyontheproblemofobjectmatchingandrecogni-tionundercontextualconstraints.AnMAP-MRFmodelisthenformulatedfollowingthesystematicapproachsummarizedinSection1.3.4.Thelabelingofasceneintermsofamodel1objectisconsideredasanMRF.TheoptimallabelingoftheMRFiso
3、btainedbyusingtheMAPprinciple.Thematchingofdifferenttypesoffeaturesandmultipleobjectsisdiscussed.Arelatedis-sue,MRFparameterestimationforobjectmatchingandrecognition,willbestudiedinChapter7.WethenderivetwoMRFmodelsforposecomputation,posemeaningthegeometrictransformationfromonecoordina
4、tesystemtoanother.Invisualmatching,thetransformationisfromthescene(image)tothemodelobjectconsidered(orviceversa).Inderivedmodels,thetransformationisfromasetofobjectfeaturestoasetofimagefeatures.TheyminimizeposteriorenergiesderivedfortheMAPposeestimation,possiblytogetherwithanMRFforma
5、tching.4.1MatchingunderRelationalConstraintsInhigh-levelimageanalysis,wearedealingwithimagefeatures,suchascriticalpoints,lines,andsurfacepatches,thataremoreabstractthanimagepixels.Suchfeaturesinascenearenotonlyattributedby(unary)propertiesaboutthefeaturesthemselvesbutalsorelatedtoeac
6、hotherbyrelations1Inthischapter,theword“model”isusedtorefertobothmathematicalvisionmodelsandobjectmodels.S.Z.Li,MarkovRandomFieldModelinginImageAnalysis,91AdvancesinPatternRecognition,DOI:10.1007/978-1-84800-279-14,cSpringer-VerlagLondonLimited2009924.High-LevelMRFModelsbetweenthem.
7、Inotherwords,anobjectorasceneisrepresentedbyfeaturesconstrainedbythepropertiesandrelations.Itisthebilateralorhigher-orderrelationsthatconveythecontextualconstraints.Theyplayacrucialroleinvisualpatternmatching.4.1.1RelationalStructureRepresentationThefeatures,propertiesandrelationscan
8、bedenotedcom
