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 > 工程科技 > 城乡/园林规划 > Multi-channel Segmental Hidden Markov Models for
侵权投诉

Multi-channel Segmental Hidden Markov Models for

多通道分割

Observationdensityfunctions:

j

p(Oji|Si=m, )=

τi

(j,i)jj

N(µ|µµ,m,Σµ,m)N(Ok|µ,Σjm)dµ,

k=1

(3)

wherep(S,F,O|Θ)isgivenin(4)andPe(Θ)istheentropic

priorofmodelΘthat,inthiswork,dependsonAas

w

w

Pe(Θ)∝expPp,qlogPp,q,(7)

w

p

q

w

wherePp,qdenotesatransitionprobabilityinAw.Accord-ingly,intheM-stepoftheEMalgorithm,wewillupdatethetransitionprobabilitiesbymaximizingtheentropicprioras,

w w

A =argmax{log(Pe(Θ|O))+λp(Pp,q 1)},

A

w

p

q

jj

where ={µjµ,m,Σµ,m,Σm|j=1,2;m=1,...,Nj}speci esthetwosegmentalmodels,Njisthenumber

(j,i)

ofstatesinchannelj(hereN1=N2=4),andOkdenotestheobservationofframekinshotiandchan-nelj.Wereferthereaderto[4]formoredetails.

Giventhedual-channelobservationsofTshots,O={Oji|i=1,...,T,j=1,2},thejointlikelihoodisde nedas:

p(S,F,O|Θ)=

1212P(S1)P(S1)P(F1|S1,S1)

T

12

P(Fi|Si,Si)

i=1

T 2

jj

P(Si|Si 1,Fi 1)

T 2

j

p(Oji|Si).

(4)

i=2j=1i=1j=1

TheExpectationMaximization(EM)algorithmisoftenused

forlearningaDBN,suchastheMCSHMM,which ndstheoptimalparametersbymaximizingthelikelihoodfunction(e.g.,(4))throughiterativeExpectation(theE-step)andMaximization(theM-step).However,themaximumlikeli-hood(ML)learningoftheMCSHMMcouldbeproblematicduetothecomplexstatespacespeci edbySandF,andalargesetofmodelparameters,suchasAgivenin(2).

3.2LearningoftheMCSHMM

Therearetwoaspectsinmodellearning,structurelearn-ingandparameterlearning.Theformeroneaimsat ndingacompactande ectivemodelstructurebysimplifyingthestatespaceandreducingtheparameterset,andthelat-teronetriestooptimizethemodelparametersgivenapre-de nedmodelstructure.Moreadvancingly,thetwolearn-ingprocessescouldbeuni edintooneframeworkwherethemodelstructureandparameterscanbeoptimizedsimultane-ously[13,2].Inthiswork,wewillusetheideasofentropicpriorandparameterextinctionproposedin[2]thatresultinamaximumaposteriori(MAP)estimator.ThisMAPestimatorcanbeincorporatedintotheM-stepoftheEMalgorithmwhichwillencourageamaximallystructuredandminimallyambiguousmodel.Thisisaccomplishedbytrim-mingtheweaklysupportedparametersandstates,leadingtoacompactmodelwithgooddeterminism.

Inthiswork,theMAPestimatorfocusesonAde nedin(2)thathasmanypossiblestatetransitionsduetoalargenumberofpossiblestatesofF(N1×N2=16here).Inotherwords,wewanttosimplifyAbykeepingonlyimportantstatetransitions,thatwoulde ectivelyreducethenumberofusefulstatesinFandbalancetheincomingandoutgo-ingtransitionsbetweenthetwolayers.Consequentially,theMAP-basedEMestimator ndtheoptimalparameterby

Θ =argmaxPe(Θ|O),

Θ

(8)

whereistheLagrangemultipliertoensure=1.Usingasimilaroptimizationtechniquediscussedin[2],theMAPestimateofΘcanbeachievedbysettingthederiva-tiveoflog-posteriorgivenin(8)tozero.Consequentially,thisstepenhancestheimportantstatesinFanddrivestheweaklysupportedonestowardszero.Accordingtothetran-sitionsofsmallprobabilities,thestatesinFthatareseldomvisitedcouldbefoundandeventuallyeliminated.Herebythestatespaceisoptimizedbybalancingthestatetransi-tionsbetweenthetwolayers.AfterAisoptimized,otherparametersinΘcanalsobeupdatedintheM-step.

Weresorttothejunctiontreealgorithmin[10]toimple-menttheMAP-basedEMalgorithm.ThejunctiontreeisanauxiliarydatastructurethatcanconvertaDirectedAcyclicGraph(DAG),suchastheMCSHMM,intoanundirectedgraphbyeliminatingcyclesinthegraph,sothatbeliefprop-agationcanbee ectivelyperformedonthemodi edgraphforinferenceandlearning.OurEMalgorithmimplementa-tionwasbasedontheBayesNetMatlabToolboxdevelopedbyKPMurphy1.Afterlearning,theViterbialgorithmcanbeusedtoobtainoptimalstatesequencesforbothchannelsatthe rstlayer,i.e.,S,whichencodestwosemanticstruc-tures,i.e.,cameraviewsandplaytypes.ThesecondlayerstatesetFisavailablethatshowsthedependencybetweentwosemanticstructuresanditistrimmedtosixstatesthatislessthan40%ofthecompletestateset.

λwp

w

qPp,q

4.EXPERIMENTALRESULTS

ThealgorithmwasimplementedinMatlab7.0andtestedonaPCwith3.2GCPUand1Gmemoryforsix30-minNFLvideos.TheproposedMCSHMMiscomparedwithourpreviousmethodsthatincludetheHMMwithGaussianemission(HMM(1))andGMMemission(HMM(2))[3],andtheembeddedGMM-HMM(HMM(3))in[5]andtheSHMMin[4].WealsoimplementedtwoCHMM-basedapproaches[1],amongwhichCHMM(1)usesshot-wisefeaturesliketheHMM(1)andCHMM(2)frame-wisefeaturesliketheSHMM.Weadoptedacoarse-to- neprogressiveinitializationstrat-egythatusesthetrainingresultofasimplermodeltoini-tializeamorecomplexmodel.Forexample,weusedtheK-meanresulttoinitializeHMM(1)whosetrainingresultwasusedtoinitializeHMM(2),andthetrainingresultofSHMMwasusedtoinitializeMCSHMM.Additionally,somepriorknowledgecanalsobeused.Forexample,the rstshotisalwaysinthecentralview.TheinitializationofAmayre-quiresomeactualcalculationofthefrequenciesofdi erentstatetransitionsinacoupleofgames.

(5)

where

Pe(Θ|O)∝P(O|Θ)Pe(Θ)=

S,F

p(S,F,O|Θ)Pe(Θ),

(6)

上一页第3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