Inthispaper,weproposeadynamicmulti-descriptorfusion(DMDF)approachtoimprovingtheretrievalaccuracyof3-dimensional(3D)modelretrievalsystems.First,anindependentretrievallistisgeneratedbyusingeachinpidualdescriptor.Second,weproposeanautomaticrelevant/irrelevantmodelsselection(ARMS)approachtoselectingtherelevantandirrelevant3Dmodelsautomaticallywithoutanyuserinteraction.Aweighteddistance,inwhichtheweightassociatedwitheachinpidualdescriptorislearntbyusingtheselectedrelevantandirrelevantmodels,isusedtomeasurethesimilaritybetweentwo3Dmodels.Furthermore,adescriptor-dependentadaptivequerypointmovement(AQPM)approachisemployedtoupdateeveryfeaturevector.Thissetofnewfeaturevectorsisusedtoindex3Dmodelsinthenextsearchprocess.Four3DmodeldatabasesareusedtocomparetheretrievalaccuracyofourproposedDMDFapproachwithseveraldescriptorsaswellassomewell-knowninformationfusionmethods.ExperimentalresultshaveshownthatourproposedDMDFapproachprovidesapromisingretrievalresultandalwaysyieldsthebestretrievalaccuracy.
