文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > 人文社科 > 文学研究 > Mining Discriminative Co-occurrence Patterns for Visual Recognition
侵权投诉

Mining Discriminative Co-occurrence Patterns for Visual Recognition

Mining Discriminative Co-occurrence Patterns for Visual Recognition

MiningDiscriminativeCo-occurrencePatternsforVisualRecognition

JunsongYuanSchoolofEEE

NanyangTechnologicalUniversity

Singapore639798

jsyuan@ntu.edu.sg

MingYang

Dept.ofMediaAnalyticsNECLaboratoriesAmericaCupertino,CA,95014USA

myang@sv.nec-labs.com

YingWuEECSDept.

NorthwesternUniversityEvanston,IL,60208USA

yingwu@eecs.northwestern.edu

Abstract

Theco-occurrencepattern,acombinationofbinaryorlocalfeatures,ismorediscriminativethanindividualfea-turesandhasshownitsadvantagesinobject,scene,andactionrecognition.Wediscusstwotypesofco-occurrencepatternsthatarecomplementarytoeachother,thecon-junction(AND)anddisjunction(OR)ofbinaryfeatures.Thenecessaryconditionofidentifyingdiscriminativeco-occurrencepatternsis rstlyprovided.Thenweproposeanoveldataminingmethodtoef cientlydiscovertheop-timalco-occurrencepatternwithminimumempiricalerror,despitethenoisytrainingdataset.ThisminingprocedureofANDandORpatternsisreadilyintegratedtoboosting,whichimprovesthegeneralizationabilityovertheconven-tionalboostingdecisiontreesandboostingdecisionstumps.Ourversatileexperimentsonobject,scene,andactioncat-egorizationvalidatetheadvantagesofthediscovereddis-criminativeco-occurrencepatterns.

Figure1.Byinferringbinaryvisualattributesfromrawimagefea-tures,suchaswheels,furry[6][4],wecandistinguishbikesfrompeoplebyaco-occurrenceofcertainattributes,forexampleabikehasmetalandwheels,butdoesnothavehead.Giventwoclassesofobjects,describedby(possiblyquitenoisy)attributes,canweef cientlydiscovertheco-occurrenceofattributesthatcanbestdiscriminatethem?

1.Introduction

Duetothecompositionalpropertyofvisualobjects,scenes,andactions,thediscoveryofdiscriminativeco-occurrencepatternisoffundamentalimportanceinrec-ognizingthem.Althoughtheextractedfeatures,suchascolor,texture,shape,ormotionfeatures,canbequiteweakindividually,anappropriatecombinationofthemwillbringastrongfeaturewhichismuchmorediscrimina-tive[31][29][3][16][2][9].Therehasbeenarecenttrendinminingco-occurrencepatternsforvisualrecogni-tion.Forexample,everyreal-worldobjectisassociatedwithnumerousvisualattributesintermsofitsmaterial,structure,shape,etc,[11][4][10],althoughitisdif culttodifferentiatethemusingasinglevisualattribute,theycanbewelldistinguishedbytheco-occurrenceofspeci cat-tributes,asillustratedinFig.1.

Inabinaryclassi cationproblem,givenacollectionofNbinaryfeatures,theproblemofco-occurrencepat-ternminingistoselectasubsetfromtheseNfea-tures,suchthattheco-occurrenceofthemcanbestdis-2777

criminatethetwoclasses.Inspiteofmanyprevi-ousworksinminingandintegratingco-occurrencepat-terns[22][20][3][29][33][16][24][9][27],noneofthesemethodsistargetedat ndingthemostdiscrimina-tiveco-occurrencepatternwiththesmallestclassi cationerror.GivenNbinaryfeatures,becausetheco-occurrencepatterncancontainanarbitrarynumberoffeatures(uptoN),thetotalnumberofcandidatesofco-occurrencepat-ternsisexponentiallylarge(e.g.3Nifconsideringtheneg-ativevalueor2Nifonlyconsideringthepositivevalue.)Asaresult,itiscomputationallyintractabletoperformanexhaustivesearch.Evenworse,unlikeconventionalfeatureselectiontask,themonotonicpropertyoffeaturesubsetdoesnotholdinsearchingco-occurrencepatterns.Namely,a(K+1)-orderbinaryfeatureisnotnecessarilybetterthanaK-orderone.Therefore,thebranch-and-boundsearchcannotbeapplieddirectly[18].Existingapproachesforco-occurrencepatternsearch,suchassequentialfor-wardselection[16][24],orrecentdatamining-drivenap-proaches[22][20][3][29][33],donotguaranteetheop-timalityoftheselectedco-occurrencepatterns.Ingeneral,whenthetrainingdataisnoisy,itisstillanopenproblemto ndtheco-occurrencepatternofthebestperformance,e.g.,minimumclassi cationerror[17].

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email