资源描述:
《[ECCV 2012] Metric Learning for Large Scale Image Classification Generalizing to New Classes at Near-Zero Cost》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、Authormanuscript,publishedin"ECCV-EuropeanConferenceonComputerVision(2012)"MetricLearningforLargeScaleImageClassification:GeneralizingtoNewClassesatNear-ZeroCostThomasMensink12,JakobVerbeek1,FlorentPerronnin2,andGabrielaCsurka21LEAR,INRIAGrenoble,655Avenuedel’Europe,38330Montbonnot,Franceemai
2、l:firstname.lastname@inria.fr2TVPA,XeroxResearchCentreEurope,6chemindeMaupertuis,38240Meylan,Franceemail:firstname.lastname@xrce.xerox.comAbstract.Weareinterestedinlarge-scaleimageclassificationandespeciallyinthesettingwhereimagescorrespondingtoneworexistingclassesarecon-tinuouslyaddedtothetr
3、ainingset.Ourgoalistodeviseclassifierswhichcanincorporatesuchimagesandclasseson-the-flyat(near)zerocost.Wecastthisproblemintooneoflearningametricwhichissharedacrossallclassesandex-plorek-nearestneighbor(k-NN)andnearestclassmean(NCM)classifiers.WelearnmetricsontheImageNet2010challengedataset,whi
4、chcontainsmorethan1.2Mtrainingimagesof1Kclasses.Surprisingly,theNCMclassifiercomparesfavorablytothemoreflexiblek-NNclassifier,andhascomparableperformancetolinearSVMs.Wealsostudythegeneralizationperformance,amongothersbyusingthelearnedmetricontheImageNet-10Kdataset,andweobtaincompeti-tiveperform
5、ance.Finally,weexplorezero-shotclassification,andshowhowthezero-shotmodelcanbecombinedveryeffectivelywithsmalltrainingdatasets.1IntroductionInthelastdecadewehavewitnessedanexplosionintheamountofimagesandvideosthataredigitallyavailable,e.g.inbroadcastingarchivesorsocialmediasharingweb-sites.On
6、lyasmallfractionofthisdataisconsistentlyannotatedandthusscalablemeth-odsareneededforannotationandretrievaltoefficientlyaccessthishugevolumeofdata.Thisneedhasbeenrecognizedinthecomputervisionresearchcommunityandlarge-hal-00722313,version1-1Aug2012scalemethodshavebecomeanactivetopicofresearchin
7、recentyears,seeamongoth-ers[1,2,3,4,5,6,7,8,9,10].TheintroductionoftheImageNetdataset[1],whichcontainsmorethan14Mmanuallylabeledimagesof22Kclasses,hasprovidedanimportantbenchmarkforlarge-scaleimageclassificationandannotationalgorithms.Inthispaperwef