資源描述:
《[2008IJCV]Robust object detection with interleaved categorization and segmentation.pdf》由會(huì)員上傳分享,免費(fèi)在線閱讀,更多相關(guān)內(nèi)容在教育資源-天天文庫(kù)。
1、ndSubmissiontotheIJCVSpecialIssueonLearningforVisionandVisionforLearning,Sept.2005,2revisedversionAug.2007.RobustObjectDetectionwithInterleavedCategorizationandSegmentationBastianLeibe1,AlesLeonardisˇ2,andBerntSchiele3Abstract—Thispaperpresentsanovelmethodfordetectingtheobjectsinthe?rstplaceandto
2、separatethemfromtheandlocalizingobjectsofavisualcategoryinclutteredreal-worldbackground.scenes.Ourapproachconsidersobjectcategorizationand?gure-Historically,thisstepof?gure-groundsegmentationhasgroundsegmentationastwointerleavedprocessesthatcloselylongbeenseenasanimportantandevennecessaryprecurso
3、rcollaboratetowardsacommongoal.Asshowninourwork,thetightcouplingbetweenthosetwoprocessesallowsthemtobene?tforobjectrecognition[45].Inthiscontext,segmentationisfromeachotherandimprovethecombinedperformance.mostlyde?nedasadatadriven,thatisbottom-up,process.Thecorepartofourapproachisahighly?exiblele
4、arnedrep-However,exceptforcaseswhereadditionalcuessuchasresentationforobjectshapethatcancombinetheinformationob-motionorstereocouldbeused,purelybottom-upapproachesservedondifferenttrainingexamplesinaprobabilisticextensionhavesofarbeenunabletoyield?gure-groundsegmentationsoftheGeneralizedHoughTran
5、sform.Theresultingapproachcandetectcategoricalobjectsinnovelimagesandautomaticallyinferofsuf?cientqualityforobjectcategorization.Thisisalsodueaprobabilisticsegmentationfromtherecognitionresult.Thistothefactthatthenotionandde?nitionofwhatconstitutesansegmentationistheninturnusedtoagainimproverecog
6、nitionobjectislargelytask-speci?candcannotbeansweredinanun-byallowingthesystemtofocusitseffortsonobjectpixelsandtoinformedway.Thegeneralfailuretoachievetask-independentdiscardmisleadingin?uencesfromthebackground.Moreover,segmentation,togetherwiththesuccessofappearance-basedtheinformationfromwhere
7、intheimageahypothesisdrawsitssupportisemployedinanMDLbasedhypothesisveri?cationmethodstoproviderecognitionresultswithoutpriorsegmenta-stagetoresolveambiguitiesbetweenoverlappinghypothesesandtion,hasledtotheseparationof