快速Mean-shift聚类
datasetclusteringpreprocessusingKD-tree,samplefeaturespacecomputing,meanshiftmodescomputinginreducedfeaturespace,modesinterpolation.Wealsopresentthecom-plexityanalysisandGPUimplementationoftheproposedalgorithm.
3.1.Reviewofmeanshift
Givenpointdataset{xi}ni=1,wherexi∈Rd
isddimensionalfeaturevector,eachisassociatedwithabandwidthvaluehi>0.Anadaptivenonparametricestimatorofthisdatasetisde nedas
2 fK(x)=1n∑n1k x xi i=1h i
hi (1)
whereK(x)=
ck,dk x 2
>0iskernelfunctionsatisfying
K(x)≥0andRdK(x)dx=1,Bytakingthegradientof(1)thefollowingexpressioncanbeobtained.
K(x)=2cn1
ng x xi 2
f m(x)(2)i∑=1hihwhereg(x)= k (x),andm(x)istheso-calledmeanshift
vector
∑n
i=11 x x 2 i=hxig
m(x)i
i ∑ni=11 x xi
x(3)hg i
i
Theexpression(3)showsthatatlocationxtheweighted
averageofthedatapointsselectedwithkernelKispropor-tionaltothenormalizeddensitygradientestimateobtainedwithkernelK.Themeanshiftvectorthuspointstowardthedirectionofmaximumincreaseinthedensity.Thefollow-inggradient-ascentprocesswithanadaptivestepsizeuntilconvergenceconstitutesthecoreofthemeanshiftclusteringprocedure
g x ∑ni=11
i xi 2x
hxi i
ii+1=,j∑n
i=11 x x=1,2,...(4)ii hg i
i
Thestartingpointoftheprocedurex
icanbechosenaspointsxi,andtheconvergencepointsoftheiterativepro-cedurearethemodesofthedensity.Theallpointsthatcon-vergetothesamemodearecollectedandconsideredasacluster.Moredetailsaredescribedin[CM02].
3.2.Fastmeanshiftcomputing
Theweightedaverageofexpression(4)isthemosttimecon-sumingcomputingofmean-shiftwhenthenumbernofthedatasetislarge(forexample,109pixelinvideostreaming).Givenanarbitrarysetofpoint{xi}ni=1withfeaturevectorof
c2010TheAuthor(s)Journalcompilation
c2010TheEurographicsAssociationandBlackwellPublishingLtd.ddimension,anaivecomputationofmeanshiftvectorex-pression(4)wouldtakeO(dn2)time,aseverypointinteractswitheveryotherpoint.Asimplewaytoacceleratethemeanshiftprocedureisusingtheweightedaverageoftheclosestpointsofxi,andthebandwidthvaluehicanbesetdepend-ingontheneighborhoodsize.However,usingthisscheme,wehavetoperformthenearestneighborhoodsearch,whichisalsoatimeconsumingoperationforlargedataset,espe-ciallyfordatasetwithhighdimensionvector.
Toacceleratetheweightedaverageoperationofexpres-sion(4),insteadofcomputingexpression(4)forindivid-ualpointsinthedataset,weapproximateoriginaldatasetbypiece-wiselinearsegmentsinthefeaturespacebasedonsimilaritymeasure,eachrepresentedbyaclusterofnearbypixels,andthesizeofeachclusterisadaptedtothesizeofthesimilarfeaturespace.Thegeneratedclusterscanbecon-sideredasthesamplesofthedataset,whichisofamuchsmallernumberthanthenumberofpoint.Theninsteadofsolvingthemeanshiftprocedure(4)directlyonindivid-ualpointsasdoneinpreviousmethods,wesolveitonthesamplesbasedonGaussianweightedaverageonaneigh-borhood, nallyweinterpolatetheclusteringresultstotheoriginaldataset.
Weclusterthepointdatabasedonsimilaritymeasurebe-tweenthepoints,whichisde nedinthefeaturespaceofinputdataset.Wede nethesimilaritybetweenthepointsusingbothspatiallocalitypandvaluevofpoint,whichcon-stitutesthefeaturespaceoftheinputdataset.Forexample,inimagecase,pointxisapixelwithitspositionp=(x,y)anditscolorvaluev=(r,g,b)(Labcolorspace).Thus,eachpointxisa ve-dimensionalfeaturevectorwhoseaxesarex=(p,v)=(x,y,r,g,b).Asstatedin[AP08][XLJ 09],thesimilaritymeasure (oraf nity)betweentwopointsde nedasz canbe
ij=exp xi xj 2
,thepositionpandvaluevalsocanbeweightedbyparameters.Forvideo,thefeaturevectorcanbeexpandedtoincludetheframeindextandmo-tionestimation ofpointx,andfeaturevectorisexpressedasseven-dimensionalvectorx=(p,v,t, ).Forimage,wecomputethemeanshiftclusteringprocedure(4)infeaturespacewhereeachpointisassociatedwithbothspatiallocal-itypandvaluev.
3.2.1.KD-Treeadaptiveclustering
WeapplyKD-treetoadaptivelyclusterthedatasetinthefeaturespace,andsubdivide nelyintheregionswherethepointfeaturevectorsaredifferent,subdividecoarselyintheregionswherethefeaturevectorsaresimilar.InimageKD-treeclustering,forexample,wesubdividethehomogeneousregionscoarsely,whilesubdividetheedgesregions nely.Thenbyrepresentingeachclusterwithasample,wecanre-ceiveanaccurateapproximatefeaturespaceoftheoriginaldatasetwithmuchlesssamples.
KD-treeclustersthedatasetbasedonfeaturespaceinatopdownway.Startingfromarootcell,thetoprectangular
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