文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > Paper number 648 Bayesian structure from motion
侵权投诉

Paper number 648 Bayesian structure from motion

Abstract:We formulate structure from motion as a Bayesian inference problem, and use a Markov chain Monte Carlo sampler to sample the posterior on this problem. This results in a method that can identify both small and large tracker errors, and yields reco

(resp.),andthenswapsvariables;thesweepscon-tinueuntilconvergence(whichisguaranteed).We

nowcomputeanaf netransformation

suchthatandisminimised;then

.Wenowdrawasam-plefromthefullconditionaloneachbitinthemaskma-trix,givenandtoobtainThestartstateisthen.

3.2Results

Whilenoconvergencediagnosticisavailable,itisclearfromthenumberofmeasurementsmarkedgoodandthequalityof tthatthesamplerhasburntin( gure4).Asam-plerstartedasdescribedaboveburnsinwithin100samples(whichcanbedrawninabout1houraMacintoshcom-puterusingMatlab).Resultsareobtainedusingthehoteldataset,courtesyoftheModelingbyVideotapinggroupintheRoboticsInstitute,CarnegieMellonUniversity.

Figure1illustratesthekindofinformationasamplercanproduce.Noticethattheestimateofvariationinin-ferredpointpositionisobtainedwithoutgeometricanaly-sis;ifadatasetisobtainedbyacameratranslatinginitsplane,thesamplerwillreturnasetofsampleswithsub-stantialvarianceperpendiculartothatplanewithoutfur-therado.Figure2illustratestrackererrorsidenti edbyasampler;theseerrorsareinthehoteldatasetasdistributed.3.2.1

Stability

Reconstructionscannotbecomparedonthebasisofaccu-racy,because“groundtruth”isnotavailable.However,wecandemonstratetheBayesianmethodissigni cantlymorestableundertrackererrorsandnoise.

Inthe40framesused,sixpointmeasurementsinnineframesareaffectedbysmalltrackererrorsasshownin g-ure2.These(verysmall)errorsaffectthereconstructionobtainedusingthefactorisationmethodbecausethefac-torisationofamatrixisafunctionofallentries.

BecausethereconstructionisinsomeunknownscaledEuclideanframe,reconstructionsarebestcomparedbycomparinganglessubtendedbycorrespondingtriplesofpoints,andbycomparingdistancesbetweencorrespond-ingpointsscaledtominimizetheerrors.

Tocomparethestabilityofthemethods,wenowin-troducelargertrackererrors;asmallpercentageofdatapoints,randomlyselected,arereplacedwithdrawsfromauniformdistributionontheimageplane.Ifthesepointsareincludedinthefactorisation,theresultsareessentiallymeaningless.Toprovideafaircomparison,weusefactori-sationsobtainedusingthemethodofsection3.1.1(thesearethestartpointsofoursampler).Thesereconstruc-tionsareguaranteedtoignorelargeerrorpointsbutwill

Paper number 648 Bayesian structure from motion

Log-posteriorvaluesforevery ftiethsampleoftheapprox-imationtodrawnbyoursampler.Thevariationinvaluessuggeststhatthesamplerexploresitsdomaineffectively.

ignoreasigni cantpercentageofthedata.Incompari-son,thesamplerquicklyaccretesallpointsconsistentwithitsmodel,andsogivessigni cantlymorestablemeasure-ments(cf[25],whichusesmaximumlikelihoodtoidentifycorrespondences).

4Bayesianmotionsegmentation

Wedescribeasegmentationmethodforthecaseofamaximumoftwoobjects;extendingthistothecaseofanunknownnumberofobjectsisstraightforward.Werequireanewvectorofdiscretevariables,oneperpoint(i.e.onepercolumnof).Thesevariablestakethreevalues,al-locatingthecolumntonoise,objectoneorobjecttwore-spectively.Writeforthenumberofcolumnsallo-catedtonoiseandforthetotalnumberofobjects(whichcanbezero,oneortwo).Thestateofthesampler

isnowgivenby

.Thelikeli-hoodisobtainedbyevaluatingequation1foreachobjectseparately.Theprioronandisasbefore.Priorsonandareasbefore.Wemultiplybyafurtherpriortermthatpenalizeslargenumbersofnoisecolumnsandsplittingasingleobjectintotwocomponents:

4.1Samplingforsegmentation

Thediscretecomponentofthedomainofthisposte-riorisextremelylarge(oftheorderof).Anap-propriatestrategypartitionsthevariables.First,wesam-pleanapproximationtothemarginalon,

上一页第6页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email