[ICCV 2011] Automatic Salient Object Extraction with Contextual Cue

《[ICCV 2011] Automatic Salient Object Extraction with Contextual Cue》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、AutomaticSalientObjectExtractionwithContextualCueLeWang,JianruXue,NanningZhengGangHuaInstituteofArtificialIntelligenceandRoboticsDepartmentofComputerScienceXi’anJiaotongUniversity,ChinaStevensInstituteofTechnology,USA{lwang,jrxue,nnzheng}@aiar.xjtu.edu.cnganghua@gmail.comAbstracts[3,19].

2、Therearemanymethodsemployingcontextcuesforobjectcategorization[15,17]andmulti-classsegmenta-Wepresentamethodforautomaticallyextractingsalien-tion[9],butfewmethodsusingcontextaimedatforegroundtobjectfromasingleimage,whichiscastinanenergyobjectsegmentationareproposed.Intuitively,theyshoul

3、dminimizationframework.Unlikemostpreviousmethodsprovidebeneficialinformationforseparatingaforegroundthatonlyleverageappearancecues,weemployanauto-objectfromitsbackground.Thismotivatedustoexplorecontextcueasacomplementarydataterm.Benefittingfromtheusageofcontextinformationforthetaskofautom

4、aticagenericsaliencymodelforbootstrapping,thesegmenta-salientobjectextractionfromasingleimage.tionofthesalientobjectandthelearningoftheauto-contextInparticular,wecasttheauto-contextmodelbyTu[20]modelareiterativelyperformedwithoutanyuserinterven-intoanenergyminimizationformulationtoimpro

5、vebothtion.Uponconvergence,weobtainnotonlyaclearsepara-theefficiencyandaccuracyforforegroundobjectsegmenta-tionofthesalientobject,butalsoanauto-contextclassifiertion.Theauto-contextmodelbuildsamulti-layerBoostingwhichcanbeusedtorecognizethesametypeofobjectinclassifieronimagefeaturesandcont

6、extfeaturessurround-otherimages.Ourexperimentsonfourbenchmarksdemon-ingapixeltopredictifthispixelisassociatedwiththetar-stratedtheefficacyoftheaddedcontextualcue.Itisshowngetconcept,wheresubsequentlayerisworkingontheclas-thatourmethodcomparesfavorablywiththestate-of-the-sificationmapsfrom

7、thepreviouslayer.Hencethroughtheart,someofwhichevenembraceduserinteractions.layeredlearningprocess,itautomaticallytakesmorespatialcontextintoconsiderationwhenclassifyingonepixel.Nevertheless,learningboththeappearancemodeland1.Introductiontheauto-contextmodeloftheforeground/back