[2008IJCV]Robust object detection with interleaved categorization and segmentation.pdf

《[2008IJCV]Robust object detection with interleaved categorization and segmentation.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、ndSubmissiontotheIJCVSpecialIssueonLearningforVisionandVisionforLearning,Sept.2005,2revisedversionAug.2007.RobustObjectDetectionwithInterleavedCategorizationandSegmentationBastianLeibe1,AlesLeonardisˇ2,andBerntSchiele3Abstract—Thispaperpresentsanovelmethodfordetectingtheobjectsinthefirstplaceandto

2、separatethemfromtheandlocalizingobjectsofavisualcategoryinclutteredreal-worldbackground.scenes.Ourapproachconsidersobjectcategorizationandfigure-Historically,thisstepoffigure-groundsegmentationhasgroundsegmentationastwointerleavedprocessesthatcloselylongbeenseenasanimportantandevennecessaryprecurso

3、rcollaboratetowardsacommongoal.Asshowninourwork,thetightcouplingbetweenthosetwoprocessesallowsthemtobenefitforobjectrecognition[45].Inthiscontext,segmentationisfromeachotherandimprovethecombinedperformance.mostlydefinedasadatadriven,thatisbottom-up,process.Thecorepartofourapproachisahighlyflexiblele

4、arnedrep-However,exceptforcaseswhereadditionalcuessuchasresentationforobjectshapethatcancombinetheinformationob-motionorstereocouldbeused,purelybottom-upapproachesservedondifferenttrainingexamplesinaprobabilisticextensionhavesofarbeenunabletoyieldfigure-groundsegmentationsoftheGeneralizedHoughTran

5、sform.Theresultingapproachcandetectcategoricalobjectsinnovelimagesandautomaticallyinferofsufficientqualityforobjectcategorization.Thisisalsodueaprobabilisticsegmentationfromtherecognitionresult.Thistothefactthatthenotionanddefinitionofwhatconstitutesansegmentationistheninturnusedtoagainimproverecog

6、nitionobjectislargelytask-specificandcannotbeansweredinanun-byallowingthesystemtofocusitseffortsonobjectpixelsandtoinformedway.Thegeneralfailuretoachievetask-independentdiscardmisleadinginfluencesfromthebackground.Moreover,segmentation,togetherwiththesuccessofappearance-basedtheinformationfromwhere

7、intheimageahypothesisdrawsitssupportisemployedinanMDLbasedhypothesisverificationmethodstoproviderecognitionresultswithoutpriorsegmenta-stagetoresolveambiguitiesbetweenoverlappinghypothesesandtion,hasledtotheseparationof