Multimodal learning for view-based 3D object classification

《Multimodal learning for view-based 3D object classification》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、Neurocomputing195(2016)23–29ContentslistsavailableatScienceDirectNeurocomputingjournalhomepage:www.elsevier.com/locate/neucomMultimodallearningforview-based3Dobjectclassificationa,ba,ba,b,nFuhaiChen,RongrongJi,LiujuanCaoaFujianKeyLaboratoryofSensingandComputingforSmartCity,ChinabSchoolofInformatio

2、nScienceandEngineering,XiamenUniversity,Xiamen,Fujian361005,ChinaarticleinfoabstractArticlehistory:Nowadaysbyemployingmanymachinelearningandpatternclassificationmethodsinobjectclassification,Received7April2015theview-based3Dobjectclassification,anemergingresearchtopic,becomesamajorresearchfocus.Rece

3、ivedinrevisedformHowever,mostexistingresearchesfocusononlyasinglemodalityofimagefeaturesfortheobjectclas-25August2015sification,althoughrecentstudieshaveshownthatdifferentkindsoffeaturesmayprovidecomplementaryAccepted19September2015informationfor3Dobjectclassification.Inthispaper,weproposethemultim

4、odalsupportvectormachineAvailableonline27February2016tocombinethreemodalitiesofimagefeatures,i.e.,Siftdescriptor,OutlineFouriertransformdescriptor,andKeywords:ZernikeMomentsdescriptortodiscriminatethemultipleclassesofobject,whereeachkernelcorrespondsView-based3Dobjecttoeachmodality.Inthisway,noto

5、nlytheindependenceofeachmodalitybutalsotheinterrelationMulti-taskfeatureselectionbetweenthemarebothtakenintoconsidered.Andwefurtheremploymulti-taskfeatureselectionviatheMultimodalSVMl2-normregularizationafterfeatureextractiontoimprovetheperformanceoffinalclassification.TheexperimentsconductedinETH-

6、80imagesetdemonstratetheeffectivenessandsuperiorityofourmethod.&2016ElsevierB.V.Allrightsreserved.1.Introductionfeaturedescriptorrepresentingtheobjectshouldbedistinguishedfromotherobjectinthedifferentclassesasfaraspossible.However,asNowadayswiththerapiddevelopmentofwebsitesandtheweknow,noneofthef

7、eaturedescriptorissuitableperfectlyforthewidespreadpopularityofmobiledevices,view-based3Dobjectclassificationofallclasses.Forexample,thecolorhistogrammayclassificationandretrievalhavebeendrawingmoreandmoreperformwelloncl