C Ja Kuo In tegrated MediaSystems Cen ter Departmen ofElectricalEngineeringSystems Univ ersit of Southern California Los Angeles CA 900892564Email shanl zhidongjongw on cc kuo sipiuscedu ABSTRA CT A new motioncomp ensated frame in terp olation sc he ID: 24873
Download Pdf The PPT/PDF document "GlobalLocal MotionCompensated rame In te..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Global/LocalMotion-CompensatedterpolationLiu,ZhidongYKimandMediaSystemsofElectricalEngineering-Systems90089-2564Email:zhidong,jongw ypesofmotionblocaretypicallyusedvidethetrajectorybeteencurrentandpreviousframes.Thepixel-basedmotion eld,whileresultsinmoreaccurateinterpolation,demandsahighcomputationalcomplexitIncontrast,theblock-based eldrequiresonlyonemotionvectorforhblochhasbeenadoptedbymostvideocompressionstandardssuchasITU-TH.263/H.263+(wheretheblocksizeis16Itprovidesacceptablevisualqualitywithacarefullydesignedalgorithm.Byemploblock-basedmotion eld,Kuoetal.[7]proposedanMCIscheme,calledthedeformableblock-basedfastMCI(DB-FMCI),whichdemonstratesagoodperformance-complexitytradeo .er,itexhibitsblockingartifactsinthemotionbackgroundarea,sincethemotionbackgroundistreatedinthesamefashionastheforeground.Thatis,onlylocalmotioncompensation(LMC)isappliedtoboththeforegroundandthebackground,whichtendstoresultinaglobalmotioninterpolationerror.ByfollowingtheframeworkofDB-FMCI[7],weinestigateanenhancedyemployingglobal/localmotioncompensationinthisresearcer,sincethealgorithmisusedasanencoder-independentpost-processingunit,weareconstrainedbythedecodedblockmotionvectorsprovidedbtheencoder.us,wefacetheuncertainyofcompression-optimizedblockmotionvectors,andhaetosettleatacompromisedsolutionfortheselectedtypeofvideo.proposedelopedmodules:moduleandthehybridmotioncompensatedframeinterpolationmodule.First,theblock-basedmotionsegmentiontechniqueisdevelopedtoiterativelydivideaframeintobackgroundandforegroundblockssothatitcanbetiedwithglobal/localmotionestimation/compensation.Especially,tocopewiththeuncertainyofcompression-optimizedblockmotionvectors,itiscombinedwiththeiny-basedHVS(humanvisualsystem)segmensegmenNext,thehybridmotioncompensatedterpolationmodulekgroundandtheforeground,respectivGlobalmotioncompensationandframeinterpolationareappliedtokgroundblockswhereeitherthe6-parameteraneorthe8-parameterperspectivemodelisusedtoreducethecomputationalcomplexityandaceperspectivecorrection.Localmotioncompensationandframeinterpolationwithlocalizedtriangularpatchmappingareusedfortheforegroundarea.Byassumingaconstantmotionvelocitbeteenneighboringframes,theproposedFMCIiscapableofinsertingasmanyframesasneededbeteendecodedframeswithoutrequiringextrabits.Inaddition,theproposedMCIschemedoesnotrequireanybitstreamsynhangeofITU-TH.263/H.263+.Itiswwhiletopointoutthatresearchonglobal/localmotionestimation/compensationhasbeentradition-allyconsideredattheencoderendtoaceahighercompressionratio,e.g.[9],[10],[11].Hadglobalestimationdoneattheencoderwiththenotionofpost-processing,itshouldprovideafurtherperformancegainsothattheresultingMCImaercomethesequencetypebarrier.ThepaperisorganizedasfolloTheoerviewoftheproposedsystemisdescribedinSection2.Thebacground/foregroundsegmentationtechniqueisdetailedinSection3.Thehybridmotioncompensation/interpolationmethodsforthebackgroundandtheforegroundarediscussedinSection4.ExperimentalresultsarepresentedinSection5.ConcludingremarksandfutureworkaregiveninSection6.ProposedTheproposedhybridmotion-compensatedframeinterpolationsystemwithperspectivecorrectiononbacblocksisshowninFig.1.Itconsistsoftomainmodules,i.e.thebackground/foregroundsegmentationmoduleandthehybridmotioncompensatedframeinterpolationmodule.Theproposedschemeworkssolelyonthedecodedbitstreambyusingablock-basedapproacItacesinterpolationresultsindependentlyoftheencoderexceptfortheuseofblockmotionvectorsprovidedbytheencoder.Theuseofaailablemotionvectorsisakeydesignfactor,sinceitfundamentallylimitsthequalityofinterpolationononehandwhilerelievingthedecoderfromtheproposedtblockmotionvectors,andusethemtoaceanregion-adaptivemotioncompensatedinterpolation.Intheproposedsystem,framesandassociatedmotionvectorsofagivensequenceare rstobtainedfromtheH.263decoder.ThesequencecontainsindependentIframesanddependentPframesalongwithmotionvforeacthebackground/foregroundsegmentationalgorithmisappliedtoapairofframestobe Figure1:Theproposedmotion-compensatedframeinterpolationsystemconsistingoftomainmodules.terpolated.Asaresult,blocksineachframecanbedividedintotypes:kgroundandforegroundblocybridmotioncompensationalgorithmisthenappliedtothetodi erenypesofblockgroundblocprocessedcompensationperspectivusingthe8-parametermotionmodel.oregroundblocksaretreatedwithtraditionallocalmotioncompensation.GMCorLMCparametersforeachblockareobtained,andskippedframesarereconstructed(i.e.terpolated)inthehybridmotion-compensatedframeinterpolationmodule.Thedetaileddesignandimplementationofthesetmoduleswillbedescribedinthefollowingsections.Withoutsegmentation,eachframeisviewedasonewholeunitsothatthemotionobjectistreatedthesameasthebacAsaresult,thevisualqualityoftheinterpolatedframeisusuallynotsatisfactoryinablocterpolationobjectivkgroundandforegroundblocksandtreatthemdi erenThesegmentationmethodproposedbyKuo[7]exploitsumanvisualsystemwherepixelsbetorkswforsequencesstationarybactreatsthemotionbackgroundtheastheforegroundobject,itcompensatesthebackgroundmotionregionlocally,whichtendstointroducenotice-ableblockingartifactsinthebackgroundarea.Here,weproposeamotionbasedsegmentationmethodthatcanobjectHVSsegmentation,theproposedsegmentationschemegreatlyacebettervisualqualityduetothereductionofbackgroundartifacts.ModelsGenerallyspeaking,motionarisinginanimagesequenceisacombinationofglobalandlocalmotionactivities.Globalmotionisusuallyintroducedbycamera'soperationand/ormot,whilelocalmotioncomesfromthetofobjectsinthescene.oseparatetheforegroundfromthebackground,itisimportanttoestimateandcompensatecamera'smotion.Algorithmsforthisseparationdi erinthemodelusedtorepresentmotionasellasthemodelparameterestimationtecosetsofpoints,()and(;:::;N arechosentocalculatethemodelparameters,where(speci esthepixelpositionofthepoinbeforecameramotand()isthecorrespondingpositionaftercameramoAmongthe2-,4-,6-and8-parametermodels,2-and4-parametermodelsusuallylackthedesiredqualitysincethe2-parametermodelmodel8-parametermodelisourchoicetorepresentthebackgroundmotion.The6-parametermodel(orcalledtheanemodel)isabletocapturethepanning,zoomingand2-DrotationalTheyareexpressedasosolvethesixunknointheaboesystem,threeindependentpointswouldbesuciencoordinatespoinequationswithpointscanbeeasilyrepresentedbTheLSsolutioncanbewrittenasInasimilarwecanobtainisa33matrixwhichisgenerallynon-singular,thecomputationalcomplexityisloourapproach,theanemodelisadoptedforbothglobalandlocalmotioncompensationapplications.enthoughitoftenfailstocapturedepthandlocaldeformations,itserveswellasacompactandfastalternativeforthewholebacorforegroundblocks,DB-FMCIreliesontheanemodelforits exibleandpoerfullocalizedmotioncompensation.eadoptthe8-parametermodel(i.e.theperspectivemodel)toestimatepanning,zooming,and3-DrotationalItcanbeexpressedas a6xi+a7yii=a3xi+a4yi+a5 eralMPEG4testsequencessuchas"Foreman"and\ContainerShip"docontainstheglobalmotionthatcanbestberepresentedbypanning,zoomingandadditional3-Drotation.us,inthispaper,the8-parametermodelisalternativelychosentocompensatetheglobalmotion.ourpointswouldbesucienttosolvetheaboesystemmodel,enhancedrobustness: TheLSsolutioncanbefoundviaItiswwhiletopointoutthat,inmostexperiments,the88matrixissingularsothatthenormaldoesproperlyDecompositionmethoddwhichcanbeedasageneralizedmatrixinersealgorithm,wasadoptedtosolvethisComparetothe6-parametermodel,the8-parametermodelprovidesperspectivecorrectionforbackgroundareasandacbettervisualqualit,whileintroducesahighercomputationalcomplexityatthesametime. decoder,tationscdecoder,objectivterpolationbetwsiterativebackground/foregroundsegmentationforpartitioningthebackgroundandforegroundblocksinthedecoded(macrobloctedbyamotionvectorassociatedwiththecentralpoinByexploitingtheglobalmotionmodelobtainedfrommotioncorrespondencebeteenthecurrentandpreviousframes,framescanbeiterativelypartitionedinkground/foregroundblocksasfolloStep1:,allblocksinthecurrentframeareconsideredasbackgroundblocThecentralpoinofallkgroundblocksaretakenas(theglobalmotionmodel.correspondingpointsinthepreviousframe,whichareobtainedfrom()andthedecodedmotionvaretakenas(The6-or8-parameterscanbecalculatedaccordinglyoranormalQCIFsequence,theframehasasizeof176144pixels.Theresulting1616blocksprovides119pairsofpoints(and(),whichareenoughfortheLScalculation.Step3:Thecalculatedglobalmotionparameters()describesanestimationofthecameramotion.Byapplyingtheglobalmotionmodelonallcentralpoints()inthebackgroundblocksoftheframe,wecanobtainasetofpoints()representsidealpositionsofthesecentralpoininthepreviousframeundertheestimatedcameramotion. modeltthemotionactivitiesofforegroundblocksandalsohasaminordi erencefromthetruebacThedi erencebeteen()and()arecalculatedintheformofthedistancebet iscomparedwithaprede nedthresholdD.IfisgreaterthanD,itislikelythatthecorrespondingblockisnotabackgroundblocksothatweremoeitfromthesetofbackgroundblocOnthecon,ifislessthanD,thecorrespondingblockisstillconsideredasabackgroundblocAfterthisprocessforallkgroundblocks,anewsetofbackgroundblockscanbeobtained.Step5:oreduceblockartifacts,weattempttoremoeisolatedbackgroundandforegroundblocThisisedbyadoptingtheclosureoperationonallbloc,eachblockhaseightdirectneighboringblocoperationboringblochangeonthetblocThe8neighboringblocksofblocarerepresentedbdenotethenberofbackgroundandforegroundblocksinthese8blocksb,respectivisaforegroundblock,ifmostofitsneighborsarebackgroundblocks,i.e.,thenwillbectoabackgroundblocisabackgroundblock,ifmostofitsneighborsareforegroundblocks,i.e.,thenwillbechangedtoaforegroundblocInallothercases,remainsuncTheclosureoperationisperformedonthewholeframeintheraster-scanorder(i.e.fromlefttorightandtoptobottom).Itisrepeateduntileithertheoperationconerges(i.e.nomorechangesontheblocks)orithesthemaximumnberofiterations.Intheexperiment,themaximumnberofiterationsissettoStep6:Aftertheclosureoperation,anupdatedsetofbackgroundblockscanbeobtained.Thecentralpoinofallbackgroundblocksaretakenas()intheglobalmotionmodel.,thecorrespondingpoininthecurrentframearetakenas(Theglobalmotionparameterscanthenbecalculatedagain.repeatSteps2-5untilthebackground/foregrounddistributionisnolongercAsaresult,weobtainthe nalsegmentationresultofbackgroundandforegroundblocThesegmentedbackgroundandforegroundblocksinanimageframeaswellastheirmotionvectorsareillus-tratedinFig.3.Comparedwithmostofothervideosegmentationtechniques,thistechniqueisrough.Itdoesnoteryaccurateobjectsegmentationduetoitsblock-basednature.Hence,acombinationofblockbasedmotiontationandpixellevelinysegmentationisadoptedheretoreducetheblockingartifactsandimprothesegmentationresult. Figure3:Illustrationofthemotionsegmentedbackground/foregroundblocksinsideaframe. Noticeableblockingartifactscanbeobservedifonlyblock-basedmotionsegmentationisusedtoidentifyfore-groundandbackgroundblocOntheotherhand,iny-basedsegmentationalonecannotdetectmotionbacgroundfromtheforegroundobject.us,weproposeasegmentationschemewhichintegratestheseapproachestovideasimplyyetaccuratesegmentationresultswhichissensitivetothebackgroundmotion.TheinybasedsegmentationmethodfollowstheframeworkproposedbyKuoetal.al.Itconsistsofthreemajorobjectoperationobjectbetgreaterthanapre-de nedsmallconsideredasobjectpixel.belongsstationarybacmorphologicalclosureoperationesmallinsidethesegmentedobject.,eac4squareismatchedwithsomepre-de nedblocksbasedonHVSandisreplacedwiththemostsimilarHVSblocTherefore,thesegmentationaccuracycanreachthe44pixellevoperatorisappliedtothesetoresults.IfapixelbelongstoamovingobjectinHVSsegmentationandaforegroundblockinmotionsegmentationatthesametime,itissaidtobeaforegroundpixel.IfitbelongstoamovingobjectinHVStationandabackgroundblockinmotionsegmentation,itisconsideredasapartofthemotionbacIfitbelongstobackgroundinbothHVSandmotionsegmentations,itissaidtobestationaryPixelsbelongingobjectpossiblyboundaryerrorscausedbblockinge ect.experimentalresultsproposedtationschemecanworkecientlyande ectivThesegmentationresultoftosuccessiveframesintheoremansequenceisshowninFig.4. Figure4:Thebackground/foregroundsegmentationresultoftheForemansequence:(a)thepreviousframe(66(b)thecurrentframe(71),(c)theHVSsegmentationresult,and(d)the nalbackground/foregroundsegmentionresult. Motion-Compensatedterpolationorhybridmotion-compensatedframeinterpolation,backgroundmotionisestimatedbyusingtheglobalmotionmodel.er,unliketheglobalmotionmodelemploedinthebackground/foregroundsegmentation,wherethetandthepreviousframesprovide()and()pointsetsmainlytodeterminethebackgroundandtheforegroundblocmotionmodeladoptedhereestimatesthecorrespondingglobalfromtheinterpolatedframetothepreviousandcurrentframesforbidirectionalinterpolation.Theseparameterswillbeusedtoreconstructbackgroundpixelsintheinterpolatedframes.Themotionvectorsofforegroundblocareestimatedbyusingthelocal6-parametermodel,inwhichasetof6-parameterswillbeappliedtotheforegroundpixelconstructioninaninterpolatedframe.Step1:Calculatethecoordinatesofkeypointsets.Letusdenotethepreviousframeb,thecurrentframeb,andtheframetobeinterpolatedbpoinblocpoincompensation.denotedbisthetotalnberofcentralpoinSomepointsareassociatedwithkgroundblockswhilewithforegroundblocSVBSVFdenotethekground/foregroundpointsets,respectivehaSVBSVFbetpointhe6-)parametermodel.oreachpoincorrespondingpointinthepreviousframe,decodedpoinSVBSVFareusedtospecifythebackgroundandforegroundpoinSVBSVFmotionvelocitfromthetframepreviousframeisconstanThisconstantmotionvelocityassumptioncanprovideareasonableestimationofthecorrespondingpoinintheinterpolatedframe,whichislinearlyinterpolatedbeteobtainSVBSVFTheconstantmotionvelocityassumptionisusedonlytoestimatethevertexpoinisneededthebackgroundandforegroundmotionestimationandcompensation.Sincelinearinterpolationistoosimple,thisassumptionshouldnottoallremainingterpolatedterpolation8-parametermodels.Step2:Applyglobal/localmotionmodelsandcalculateglobal/localmotionparameters.Withthecorrespondingcoordinatesofpointsbeteentheinterpolatedframeandprevious/currentframe,wareablecalculatebothglobalandlocalmotionparametersandbuildpropermotionmodels.ormotionobject,foregroundregion,localized6-parameteranemodeltriangularpatcmethodproposedininisadopted.Withthisapproach,eachsetborhoodtralpoinhformasquareisdividedintototrianglesbyusingtheshorterdiagonalcriterion.Threeverticesoftriangularprovideauniquesolutiontoalocalized6-parametermodel,whichisusedtogenerateallpixelsinalocalregionintheinterpolatedframe.Both6-parameterand8-parametermodelscanbeappliedtoestimatetheglobalmotionbasedondi erentuseretheforegroundmotionmodel,hislocalized,allbackgroundpixelsfollowthesamemotionisrepresentedboneglobalmotionmodel.8-parameterglobalmotionmodelscanbeusingtheleast-squareestimationsolutionasdiscussedbefore.Step3:Constructinterpolatedframeswithglobal/localmotioncompensation. eusebidirectionalinterpolationsothateachpixelintheinterpolatedframeisconstructedbythecomofbacardandforwardinterpolation.Considerapixelintheinterpolatedframe.Withthebacardapproacitscorrespondinglocationinthepreviousframeiscalculatedanditsvalueisobtained.Withtheforwardapproaccorrespondingpixellocationaccordingtobetterpolatedtframeaddedtodeterminethecorrespondingpixelvalueintheinterpolatedframe.Dependingonwhetherapixelbelongstoabackgroundoraforegroundblock,motionparametersusedtointerpolateitsvaluearedi erenSinceallbackgroundblocksundergothesamemotion,onlyonesetof6-or8-parametersisneededtorepresenthisbackgroundmotion.Ifapixelbelongstothebackground,itsbacardinterpolatedvaluecanbedeterminedythebacardmotioncompensatedinterpolation,wheretheglobalmotionparametersarecalculatedbSVBSVBSVBSVBcanbeusedtocalculatethe6or8-parameterstore ectallbackgroundblocks'motionbeteentheinterpolatedframeandthecurrentframe,whichareusedtocalculatetheforwardinterpolatedpixelbelongsblocblocSincedi erentforegroundblockshaedi erentmotions,thelocalizedforegroundmotionparametersaregenerallynotthesame.Therefore,thebacardandforwardestimatedvalueoftheinterpolatedpixelhastobecalculatedaccordingtothelocalmotionmodelassociatedwiththetriangularpatchitbelongsto.ormoredetails,wereferto[7].Experimenorperformanceevaluation,wehaeappliedtheproposedschemetototestsequencesoftheQCIFformat:MissAmericaandFEspeciallythevisualqualityisextensivelycomparedthatofDB-FMCItoverifytheadditionalgainduetoextraglobal/localhybridprocessing.TheresultsareshowninFig.5andFig.6,respectivIntheMissAmericaQCIFvideosequence,theoriginalframerateofinputsequenceis30framepersecond(fps)intheencoderend,thebasicmode(i.e.nooptionalmodeisactivated)isselected,andthequantizationstep20andframeskip10areused.Therequiredbandwidthfortheencodedbitstreamisonly8kbpsduetotheadoptionofalargeframeskip.Asaresult,fromthedecoderend,wecanobtainaframerateof3fps.9framesareinsertedbeteentoadjacentframesinthedecodedbitstreamtorecoertheframeratebackto30fps.Fig.5(a)showstheframeintheoriginalsequence,(e)showsthe71frameintheoriginalsequence.ecomparetheinterpolatedframesfromthe61andthe71framesbyusingdi erentframeinterpolationmethodsinthesame gure,proposedtheoriginal66framein(c).romthem,eseethat(b),(d)arealmostthesame,novisualdi erencesisThisisduetothefactthatthebackgroundmotionintheMissAmericasequenceistoosmalltobeus,eventhoughweapplytheglobalmotioncompensationtothebackground,itisnotnoticeable.IntheForemanQCIFsequence,theoriginalframerateoftheinputsequenceisalso30framepersecond(fps)intheencoderend,thebasicmode(i.e.nooptionalmodeisactivated)isselected,andthequantizationstep13andframeskip4areused.Therequiredbandwidthfortheencodedbitstreamis34.36kbps.romthedecoderend,ecanobtainaframerateof5.93fps.4framesareinsertedbeteentoadjacentframesinthedecodedbitstreamertheframeratebackto30fps.Fig.6wsthe66frameintheoriginalsequence,wstheframeintheoriginalsequence.eshowtheinterpolatedframesfromtheandthe71framesbyusingdi erentframeinterpolationmethodsinthesame gure,where(b)isobtainedbyDB-FMCI,(d)isobtainedusingtheproposedscheme,and(c)istheoriginal66orthiscase,weseeacleardi erenceinthekgroundareain(b),(c)and(d).WiththeDB-FMCIapproach,thebackgroundisblurred,whiletheblurredregionsareremoedintheproposedschemewiththeaneglobalmotioncompensation.(d)hasasigni cantvisualer(b).botheasgoodoriginalframe.Somebackgroundmotionisstillnotcompensatedwellandrequiresfurtherpostprocessing.ThisisstillunderourtresearchanddevByreplacinganeGMCwithperspectiveGMC,visualenhancementinbackgroundareascanbeacedat Figure5:Comparisonofdi erentinterpolationmethodsonMissAmericasequence,where(a),(c),(e)showthe,the66,andthe71framesintheoriginalsequence,respectiv,and(b),(d)showthe66frameobtainedyinterpolatingbeteenthe61andthe71framesbyusingDB-FMCIandtheproposedscheme,respectivthecostofahighercomputationalcost.AsshowninFig.7,wehaetheinterpolated95framesfromaneGMCandperspectiveGMC.ThebackgroundregioninterpolatedbyperspectiveGMClooksbetterthanthatofaneGMCandinconsistencyoccurringintheupperrightfaceareaintheaneGMCframeis xedbytheperspectivGMCmethod.proposedmotion-compensatedterpolationperspectivproposedmodules,ground/foregroundsegmenmodulecompensatedterpolationmodule.foregroundcanbesegmentedfrombothstillandmotionbackgroundecientlyinthesegmentationmodule.arethentreatedwithdi erentmotionmodelstoacethemotion-compensatedframeinterpolation.Thebacgroundmotionisrepresentedbya6-or8-parameterglobalmotionmodeleperspectivecorrectionatawcomputationalcomplexit,whileforegroundmotionsaremodeledbyusinglocalizedane6-parameters.setofglobalmotionparametersisappliedtoallbackgroundblocksforinterpolation,whileeachforegroundblocusesitsownmotionparameterstoconstructtheinterpolatedpixelsinthecorrespondingtriangularpatcExperi- Figure6:Comparisonofdi erentinterpolationmethodsappliedtotheForemansequence,where(a),(c),(e)shorespectivterpolatingbetproposedrespectivtsshowthattheproposedschemeworksbetterinthemotionbackgroundareacomparedthatoftheDB-FMCImethod[7].Asigni cantvisualqualityenhancementcanbeacedtoaoidblockingartifacts.orthesmallorzeromotionbackground,thereisnonoticeablevisualdi erencebeteenthetoscInthemotioncompensationmodule,weadoptaconstantmotionvelocityconstrainttoestimatethekeypoinlocationsintheinterpolatedframe.Theyareusedinboththe8-andthe6-parametermodels.er,itmanotbealwystruethatthemotionisinconstanelocitErrorsgeneratedinkeypointscannotberecoeredinthelatermodelestimationprocedure.eareworkingonanimproedmethod,whichwillconsiderthenonlinearterpolationfromthecurrentframetothepreviousframeratherthanthelinearinterpolationofmotionvTheperformanceontheforegroundareashouldalsobeimproedbyintroducing8-parametermodeltoforegroundmotioncompensationatahighercomputationalcost.ost.A.M.TDigitalvideopr,Prenticehall,NJ,1995. ComparisonofterpolationmethodsterpolatedframebyaneGMCand(b)theinterpolated95framebyperspectiveGMC..A.N.NetravaliandJ.D.Robbins,\Motion-adaptiveinterpolationdingSymp,12.1,June,1981.1981.H.G.Musmann,P.PirschandH.J.Grallert,\Advancesinpicturecoding",c.IEEEol.73,pp.523-548,Apr.1985.1985.T.Chen,Y.Wang,H.P.Graf,andC.Swain,\Anewframeinterpolationschemefortalkingheadsequences",c.ICIP-95ol.2,pp.591-594,1995.995.R.ThomasandM.Bierling,\MotioncompensatingterpolationSignalPrImageCommunic,pp.191-212,1989.1989.C.K.WongandO.C.Au,\Fastmotioncompensatedtemporalinterpolationforvideo",inc.SPIEVisualationsandImagePressing`95,pp.1108-1118,May1995.995.T.KuoandC.-C.JayKuo,\Motion-compensatedinterpolationforlow-bit-ratevideoqualityenhancemenc.SPIEApplicofDigitalImageProl.3460,pp.277-288,1998.1998.S.F.WuandJ.Kittler,\Adi erentialmethodforsimultaneousestimationofrotation,changeofscaleandSignalPrImageCommunicol.2,no.1,pp.69-80,May1990..F.Moscheni,F.DufauxandM.Kunt,\Anewtwo-stageglobal/localground/foregroundsegmenc.ICASSP-95ol.4,pp.2261-2264,1995.1995.HirohisaJozaa,KazutoKamikura,AtsushiSagata,HiroshiKoteraandHiroshiWatanabe,\Two-StageMo-CompensationLocalSystemsforVideol.7,no.1,pp.75-85,Feb.1997.1997.Chung-TaoChu,DimitrisAnastassiouandShih-FuChang,\HierarchicalGlobalMotionEstima-tion/CompensationinLowBitrateVideoCoding",inonCircuitsand,June1997.997.W.H.Press,B.P.Flannery,S.A.TolskyandW.T.ValresinC:theartofscienti c,pp.60-71,CambridgeUnivyPress,1993.