资源描述：
《Unsupervised Visual Object Class Recognition》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。
1、UnsupervisedVisualObjectClassRecognitionA.NoulasB.J.A.Krose¨IntelligentSystemsLabAmsterdam,UniversityofAmsterdam,Kruislaan403,1098SJAmsterdam,TheNetherlandsanoulas@science.uva.nlkrose@science.uva.nlKeywords:ASCI,AnnualConference,VisualVocabulary,UnsupervisedMethodsAbstractofanobject,assumingth
2、eyareextractedrobustly,areinvarianttodifferencesinlightingcondition.TheObjectclassrecognitionisaverywellstudiedmostpromisingresultscomefromlocalimagefea-probleminthedomainofComputerVision.Lately,tures,whichareexpectedtobeinvarianttoallkindsthemostsuccessfulapproachesextractinformativeofaffinetr
3、ansformations.Themostpopularchoiceisdescriptorsfromtheimages,organizethemandusetheScaleInvariantFeatureTransform(SIFT)descrip-machinelearningtechniquesforlearningandclassifi-torspresentedin[1].cationtasks.Acommonapproachinvolvesthecon-structionofavisualvocabularywhichisusedtoex-pressthecontento
4、ftheimage,andtheapplicationoftextretrievaltechniquesforinferencepurposes.Inthispaperwepresentanunsupervisedimplementa-tionofthisapproach.1IntroductionVisualrecognitionproblemscanbedividedintotwomaincategories,objectrecognitionandobjectclassrecognition.Theformerinvolvestheefficientdetectionofasp
5、ecificobject,withcertainvisualat-tributes,forinstanceaspecificcar.Inthelattercate-gorywearenotlookingforaspecificinstancewithinaclass,butwerathertrytolearnaninternalmodelFigure1:OverviewofVisualRecognitionofObjectsthatcorrespondstoalltheobjectsofaclass.Thecon-andObjectClassesUsingAffineInvariantFe
6、aturesceptthatlabelsanobjectclasscanbeaswideorre-strictedasneeded,withconceptslikesportcars,ve-Thevectorthatdescribestheaffineinvariantfea-hicles,publicmeansoftransportationtobepossibleturesofaregion,iscalledtheaffineinvariantfea-examples.turedescriptor.ThesedescriptorsarecomputedforVisualrecogn
7、itionisaveryhardproblemsincetheselectedimagepatcheswhichcanbedetectedusingpixelvaluesthatcorrespondtomultiplepicturesofappropriatealgorithms.Manyaffineinvariantregionanobjectsuffergreattransformations.Thesetrans-detectorsexist,withacompa
