快速Mean-shift聚类
cellrepresentsallpointsinthedataset,werecursivelysplitacelltotwochildcellsadaptivelyalongadimensionthatisal-ternatedatsuccessivetreelevels.Similarto[AGDL09],eachinnercellofthetreeTrepresentsad-dimensionalrectangu-larcellwhichstoressixvariables:thedimensiondalongwhichthetreecuts,thecutvalueTcutonthatdimension,theboundsofthecellinthatdimensionTminandTmax,andpointerstoitschildrenTleftandTright.Leafnodescontainonlyad-dimensionalpointwhichisconsideredasasample.Thissamplerepresentsthepointsthatcontainedinthecell,thekd-treestoresmsamples y jm
leafj=1oforiginaldataset in md-dimensions,onesamplepointperleaf.Thesamplesyjj=1constructthereducedfeaturespaceoftheoriginaldataset.
AsillustratedinFigure1,weadaptivelyclustertheim-ageintoclusters.Theimageisadaptivelyclustered,whereattheedgeregions,moresamplesareplaced,whileattheho-mogeneousregions,coarsesamplesareplaced.Thesamplefactorcanbechangedbythestoppingcriteria.Wepresenttwothresholdsforstoppingcriteria,oneisthesizeofthecell,otheroneisthevarianceσofthesimilaritymeasurezijbetweenthepointsinthecell.Byusingthesetwothresh-oldswecangeneratedifferentsamplingfactorforimageaswellasrespectingforthefeaturedistributionofthedataset.Figure1showssomeclusterresultswithdifferentsamplingfactor.
Figure1:ImagesamplingusingadaptiveKD-tree.(a)Orig-inalimage(636×844),(b)imageclusteringwith1,050sam-ples,(c)imageclusteringwith11,305samples.
3.2.2.SamplefeatureWeobtainthesamples spacey computing
jm
j=1ind-dimensions,whichareadaptivesamples thefeaturespaceoforiginaldatay ofset.Thesamplesjm
j=1canbeconsideredaapproximatefea-turespaceoforiginalspace.Tomakeamoreaccurateap-proximateples y jmfeaturespace,wescatterthepointxitothesam-j=1basedontheaf nitysimilarityzijbetweenthepointxiandsampleyj,andobtainanaf nitysimilaritybasedsamplespace.
Similartothesplattingstagesin[AGDL09],wescattereach y pointxitoitsnearestneighborhoodN(xi)insamplesjm
j=1(WeapplytheKD-treetosearchthehighdimen-sionnearestneighborhoodN(xi)forpointxi),andcompute
anaf nitybasedsampley jforeachsampleyi.Theaf nity
basedsampley
jcanbeconsideredastheweightedsimilar-ityaverageforthosefeaturevector{xi}thatismostsimilartothesampleyi.Wecomputeandsumtheaf nitysimilar-ityzijbetween xiandeachyj∈N(xi)toobtaintheaf nity
basedsampleyjofthesampleyi:y j=y j+zijxi,theny
jisnormalizedbythesumofthesimilarityzij.Thegeneratedy jisafeaturevectorofd-dimensions.Theaf nitybasedsam-ple{y j}m
j=1isamoreaccuratesamplesoftheoriginalpointset,andwillbeusedinthemeanshiftclusteringprocedureandmodesinterpolation.Thenforeachsample,westoretwovectors,one isfeaturevectoryi,theotheristheaf nitybasedsampleyj.3.2.3.Mean-shiftmodescomputing
Afterobtainingtheaf nitybasedsamples{y j}mj=1fororig-inaldatasetfeaturespace,insteadofcomputingthemeanshiftclusteringprocedureintheoriginalspace,wecom-pute themeanshiftprocedureonthereducedfeaturespace
{yj}mj=1.Noteforeachiteration,thenearestneighborhoodN we ndforeachsampley jy
jinthesamplespace,andperformfollowinggradient-ascentprocesswithanadaptivestepsizeuntilconvergence:
∑
y uj y 2 ui∈N(y j)
yigj+1=2,j=1,2,...∑ uj y
(5)
y∈N(y )g i
j
Iteratinganin nitenumberoftimestheexpression(5)isguaranteedtobringujtothemodeinreducedfeaturespace.Practically,intheimplementation,we ndthatthemean-shiftclusteringprocedure(5)tendstoconvergeinaverysmallnumberofsteps,typicallyaround6.
Asthenumberofthesamples y m
jj=1ismuchsmaller
thanthenumberofthepoints{xi}ni=1,thatism n,thecomputationalcomplexityofmeanshiftvectorcomputinghasbeenreducedsigni cantly.Furthermore,themeanshiftisperformedintheaf nitysimilaritybasedreducedfeaturespace,whichleadstomoreaccuratemodes.WeapplytheKD-treetoperformthehighdimensionnearestneighborre-searchforeachiterationcomputing. Afterperformingmeanshiftiterationsforeachsampleyj,we{zsthereducedfeaturespace receiveymthemodesk}k=1of jj=1,whicharetheapproximatemodesoforiginaldataset.Allsamplescon-vergingtothesamemodeareclusteredtogether.
3.2.4.Modesinterpolation
Tointerpolatethemodescomputedinthereducedfeaturespacetotheoriginaldataset,onenaïveapproachisto ndanearestmodezl∈{zk}sk=1foreachpointxi.Thiscanbeconsideredasahardclustering.Asanalternativemethod,wegiveasoftclusteringmethodwhichgeneratesmoresmoothclusteringresults.Thismethodworksbyapplyingweightedbasedinterpolation.
c2010TheAuthor(s)Journalcompilation
c2010TheEurographicsAssociationandBlackwellPublishingLtd.
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