快速Mean-shift聚类
Themodeinterpolationworksasfollows,xsamplesN(x i,we ndthenearesti)iny m
foreachpointjj=1.Eachsam-
pley j∈N(xi)convergestoamodeuj,thatis,yj→uj.Basedontheaf nitysimilarityzijbetweenxiandy j,bynor-malizingtheweightszij:∑jzij=1,the nalmodearecom-putedas:xi→∑j=1zijuj.Whenwecomputemodeinterpolationoverallthesamples y them
weighted
jj=1,similarto[FK09],wewillreceivethecartoon-likeclusteringresults.Notethatintheinterpolationstage,thesizeoftheN(xi)isnotalwaysthesameasthatperformedinthesamplespacecomputingstage.Inourexperiment,wesetneighborhoodsizebetween10and20,andreceivesatis edresults.Asthesamplesaresigni cantlysmallerthanthenumberoftheoriginalpointset,usingtheGPUacceleratedKD-treenear-estneighborhoodsearch,thesearchperformingisfast.Inaddition,sincewedeterminethe nalmodeforxibasedontheweightedsimilarity,theresultsaremoreaccurate.Figure3showstheresultsusingthetwodifferentmodeselectionmethods,oneisthenearest-samplebasedmodeselection,theotherisweightedmodesinterpolation.AsillustratedinFigure3,usingtheweightedmodesinterpolation,wereceivesmootherandmoreaccurate


results.
(a)(b)(c)
Figure2:Imagesegmentationresultscomparison.(a)Theoriginalimage,(b)theresultusingnearestsamplebasedmodeselection,(c)theresultusingweightedmodesinter-polation.
3.3.ComplexityanalysisandGPUimplementationOuralgorithmacceleratesthemeanshiftproceduresbycom-putingalowerresolutionfeaturespaceandtheninterpo-latingtheclusteringresulttotheoriginaldataset.Byus-ingaKD-treestructure,weconstructareducedfeaturespaceforinputnd-dimensionalturevector:{x i}n
i=1→y mdatapointswithmfea-jj=1,andm n.AssumingthedepthofGaussiantreeisO(logm),thecomplexityoftreeconstructionisO(ndlogm).Performingnearestneigh-borhoodforeachoftheninputpointstoscattervaluesintothetreetakesO(n(logm+d))time.Ifperformingτiter-ationsformeanshiftclusteringprocedure,computingthemeanshiftiterationinthereducedspacewithmfeaturevectortakesO(τm(logm+d))time.Inthelaststage,we
c2010TheAuthor(s)Journalcompilation
c2010TheEurographicsAssociationandBlackwellPublishingLtd.up-sampleclusteringresultstotheoriginaldatasetswhichtakesO(n(logm+d))time.Recallthatm n,thisresultsinatotalcomplexityO(dnlogn).ComparedwithstandardmeanshiftprocedurewithcomplexityO(τdn2),thepro-posedmethodsigni cantlyaccelerated.
Applyingthemethodpresentedby[ZHWG08],aKD-treeisef cientlybuiltontheGPUfortheinputpointswithhigh-dimensionalfeaturevector.Threestagesofourproposedal-gorithm,includingthepointsscattering,meanshiftcluster-ingprocedureinthereducedspace,andmodesinterpola-tion,allincorporatethehighdimensionalnearestneighborsearch.AsthehighdimensionalnearestneighborsearchcanbeimplementedinparallelusingGPU[AGDL09],thus,ourmethodisevenfasttoprocesslargedatasetwithhighdi-mensionalfeaturevector.Weimplementtheproposedalgo-rithmsinCUDA,andrunitonanNVIDIAGeForceGTX285(1GB)graphicscard.Weobserveatypicalspeedupof20xoveroursingle-threadedCPUimplementationrunningonanPentium(R)Dual-CoreCPUE5200@2.50GHzwith2GBRAM,whichallowsourmethodtobeappliedinaninteractivemeanshiftclusteringframeworkformoderate-sizeddataset.
4.Applicationsandcomparisons
Weapplytheproposedfastmeanshiftclusteringtofol-lowingapplications,imagesegmentation,videosegmenta-tion,geometrymodelsegmentation,andanimatedobjectsegmentation.Wealsopresentthecomparisonresultsonbothperformanceandsegmentationqualitywiththemostrelatedmethods.OurapproachisimplementedusingC++onamachineequippedwithPentium(R)Dual-CoreCPUE5200@2.50GHzwith2GBRAM.TheGPUaccelerationisbasedonCUDA[http:doc.guandang.net]andrunonaNVIDIAGeForceGTX285(1GB)graphicscard.
4.1.Imagesegmentation
Weapplytheproposedfastmeanshiftmethodforimage
segmentation.Allpixelsthatconvergetosamemodearecollectedtogetherandareconsideredtobethesameseg-ment.Inimagecase,wede nethefeaturevectorofpixelxi=(σppi,σcci)comprisingitspositionp=(x,y)anditscolorvaluec=(r,g,b)(Labspace)whichcanbeweightedbyparametersσpandσc.Thus,eachpixelxiisa ve-dimensionalvector.
Figure3presentsthesegmentationresultsgeneratedap-plyingourfastmeanshiftmethodbasedondifferentsam-plingfactor.AsillustratedinFigure3,thereare6×106pixelsintheoriginalimage.Evenwithveryhighsamplingfactorsuchasn/m=4,096,thesegmentationresultsisstillpleasing.Withmuchlesssamples,theimagecanbeclus-teredinhighspeedevenwithoutusingGPUacceleration.It
Answers Final Exam CHEM 3080 April 9, 2012, 2 pm-5 pm CB 121 Name ... no additional paper, but you are allowed to use the back of the pages...
Academic paper Final Paper Literature review ? including the backgrand, previous research limitation, purpose, method, findings and results, futhur research, ...
final_draft_其它语言学习_外语学习_教育专区。The Final Draft Reflecting on ...8 ? by 11 inch white paper ? double spaced, size 12 font ? one inch...
120703_UNIQLO PR Plan CN final-14时 (NXPowerLite)_销售/营销_经管营销_...paper (release) ● 媒体/政府关系 ● 准备并进行面对媒体的准备 ● 紧急时...
Appendix Final exame paper of systemic anatomy for foreign studnts in WZMC Part One: 5.5 Years MBBS Major, 1 Semester of 2006-2007, Systemic Anatomy...
Tegra_Multiprocessor_Architecture_white_paper_Final_v1.1 tegra2 nvidiategra2 nvidia隐藏>> Whitepaper NVIDIA Tegra Multi-processor Architecture V1.0 P age...
1000-1220(2010)07- An approach for Building Complex Network of Web Service...In this paper, the approach of building complex network of web service is...
English_Final_Examination_Paper_A_英语考试_外语学习_教育专区。山东建筑大学往年期末及答案College English Achievement Test (Band I ) 试卷类型 A Paper One ...
take-home paper final_英语学习_外语学习_教育专区 暂无评价0人阅读0次下载举报文档 take-home paper final_英语学习_外语学习_教育专区。Yintong Liu ...
西门子面试题 A6_Answers_to_paper_testing_Final_IT/计算机_专业资料。A6答案Answer to paper testing of SCM Excellence Graduate Program 2011 I. 翻译 1. ...
热门文档