@[CVPR 2012] Weak Attributes for Large-Scale Image Retrieval

《@[CVPR 2012] Weak Attributes for Large-Scale Image Retrieval》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、WeakAttributesforLarge-ScaleImageRetrieval∗FelixX.Yuy,RongrongJiy,Ming-HenTsaix,GuangnanYey,Shih-FuChangyColumbiaUniversity,NewYork,NY10027y{yuxinnan,rrji,yegn,sfchang}@ee.columbia.eduxminghen@cs.columbia.eduAbstracttrieveimagesofabird,onecoulddescribethephysicaltraitsoffeather,beak,andbodyetc.Th

2、etaskistoretrieveim-Attribute-basedqueryoffersanintuitivewayofimageagescontainingallofthequeryattributes.Weassumeonlyretrieval,inwhichuserscandescribetheintendedsearchasmallportionofthedatabasehavethequeryattributesla-targetswithunderstandableattributes.Inthispaper,wede-beledbeforehand,andyetourg

3、oalistosearchtheentirevelopageneralandpowerfulframeworktosolvethisprob-large-scaleimagecorpus.lembyleveragingalargepoolofweakattributescomprisedAstraightforwardsolutionfortheaboveproblemistoofautomaticclassifierscoresorothermid-levelrepresenta-buildclassifiersforthequeryattributesofinterest,andtion

4、sthatcanbeeasilyacquiredwithlittleornohumansumtheindependentclassifierscorestoanswersuchmulti-labor.Weextendtheexistingretrievalmodelofmodelingattributequeries[12].Apromisingalternative,asshownindependencywithinqueryattributestomodelingdependency[23],istoanalyzethedependenciesamongqueryattributeso

5、fqueryattributesonalargepoolofweakattributes,whichandleveragesuchmulti-attributeinterdependencetomiti-ismoreexpressiveandscalable.Toefficientlylearnsuchagatethenoisesexpectedfromtheimperfectautomaticclas-largedependencymodelwithoutoverfitting,wefurtherpro-sifiersandtherebyachieverobustqueryperforman

6、ce.Anil-poseasemi-supervisedgraphicalmodeltomapeachmulti-lustrativeexampleoftheabovedependencymodelisshownattributequerytoasubsetofweakattributes.ThroughinFigure1.extensiveexperimentsoverseveralattributebenchmarks,However,[23]reliedonlyonthepre-labeledqueryat-wedemonstrateconsistentandsignificantp

7、erformanceim-tributestodesignthedependencymodel,limitingitsper-provementsoverthestate-of-the-arttechniques.Inaddi-formanceandscalability.Ononehand,userlabelingistion,wecompilethelargestmulti-attributeimageretrievalabur