Video Transmission in 3G Telephony

Embed Size (px)

Citation preview

  • 8/2/2019 Video Transmission in 3G Telephony

    1/6

    AbstractVideotransmissionin3Gwirelessenvironments

    isachallengingtaskcallingforhighcompressionefficiencyaswell as a network friendly design. Theseare themaingoals of the new ITU-TH.26L standardization effort ad-dressingconversational(i.e., videotelephony)andnon-conversational(i.e.,storage,broadcast,orstreaming)appli-

    cations. The video compression performance of H.26Ls

    VideoCodingLayertypicallyyieldsafactoroftwoormoreinbit-ratesavingswhencomparingagainstallpreviousin-ternationalvideocodingstandardsandthereforeprovidesasignificant improvement towards this end. The network-

    friendlydesigngoaloftheH.26LprojectisaddressedviatheNetworkAdaptationLayerthatisbeingdevelopedtotrans-

    portthecodedvideodataoverexistingandfuturenetworkssuchas circuit-switchedwirednetworks, IPnetworkswithRTPpacketization,and3Gwirelesssystems.Theerrorresil-

    iencefeaturesandappropriatetestmodelextensionsaten-coderanddecoderareintroducedin thispaper.Addition-ally,selectedresultsshowingthepotentialsofthesefeaturesarepresented.

    I. INTRODUCTION

    H.26L[1]isthecurrentprojectoftheITU-TVideoCod-

    ingExpertsGroup(VCEG)agroupofficiallychartered

    as ITU-T Study Group 16 Question 6. Just recently, a

    JointVideoTeam(JVT)wasformedconsistingofVCEG

    andMPEG(ISO/IECJTC1/SC29/WG11:MovingPic-

    tureExpertsGroup).ThecharteroftheJVTistofinalizetheH.26LprojectofVCEGastechnicallyalignedITU-T

    RecommendationandISOStandardcalledJVTCoding.The primary goals of theJVT projectare improved

    codingefficiency,improvednetworkadaptationandsim-

    ple syntax specification. The syntax of JVT Coding

    should permit an average reduction in bit rate by 50%

    comparedtoallpreviousstandardsforasimilardegreeof

    encoder optimization.Recentresultsshowthat thisper-

    formanceisalmostachieved[3].ThismakesJVTCoding

    anattractivecandidateforwirelessvideotransmission,as

    the resourcebit-rate is extremely costly inmobileenvi-ronments.However,toallowtransmissioninmobileenvi-

    ronmentsinadditiontocodingefficiency,anetworkad-aptationlayeranderrorresiliencefeaturesareveryimpor-

    tant.Relatingissuesexaminedseriouslyforthefirsttime

    intheH.263andMPEG-4projects[2],[4]arebeingtaken

    further in JVT Coding. The scenarios emphasized are

    primarilyforInternet,LAN,andthird-generationmobile

    wireless channels.Finally, thedesignofJVTCoding is

    strongly intended tolead toa simple and cleansolution

    avoiding any excessive quantity of optional features or

    profileconfigurations.

    Thispaperisorganizedasfollows.Wewillbrieflyin-

    troduce video transmission in 3Gnetworks andpresent

    the test conditions used in the JVT Coding project to

    evaluate rate-distortion performance.Thenwewill pro-

    videa shortoverviewofJVTCodingwithprimaryfocus

    ontheerrorresiliencefeatures.Encoderanddecodertest

    modelextensionsarediscussedthatallowtheuseoftheprovidedfeaturesinanoptimizedway.Selectedsimula-

    tion results arepresented and discussed and somecon-

    cludingremarksareprovided.

    II. VIDEOTRANSMISSIONIN3GNETWORKS

    A. Overview

    Videotransmissionformobileterminalswillbeamajor

    applicationintheupcoming3Gsystemsandmaybeakey

    factor for their success. Thedisplay of videoonmobiledevicespavestheroadtoseveralnewapplications.Three

    major service categories areidentified in theJVTstan-

    dardization process [5]: 1) conversational services for

    video telephonyandvideo conferencing,2) liveorpre-

    recordedvideostreamingservices,and3)videoinmulti-

    mediamessagingservices(MMS).

    In general, mobile devices are hand-held and con-strainedinprocessingpowerandstoragecapacity.There-

    fore,amobilevideocodecdesignmustminimizeterminal

    complexitywhileremainingconsistentwiththeefficiency

    androbustnessgoalsofthedesign.Inaddition,themobileenvironmentischaracterizedbyharshtransmissioncondi-

    tionsintermsoffadingandmulti-userinterference,which

    resultsintime-andlocation-varyingchannelconditions.

    Manyhighlysophisticatedradiolinkfeatureslikebroad-

    bandaccess,diversity techniques,fastpowercontrol,in-

    terleaving,forwarderrorcorrectionbyTurbocodes,etc.,

    are used in 3Gsystems to reduce thechannel varianceand,therefore,thebiterrorrateandradioblocklossrate.

    Entirely error-free transmission of radio blocks isa

    generally unrealistic assumption although with RLCretransmission methods, delay insensitive applicationslikeMMScanbedeliverederror-freetothemobileuser.

    In contrast, conversational and streaming serviceswith

    real-timedelayandjitterconstraintsallowforonlyavery

    limitednumberofretransmissions,ifany.Inaddition,ina

    cellularmulti-userenvironmentthetransmissioncapacity

    withineachcellislimited.Therefore,ifnewusersenter

    JVT/H.26LVIDEOTRANSMISSIONIN3GWIRELESSENVIRONMENTS

    THOMASSTOCKHAMMER,TOBIASOELBAUM

    InstituteforCommunicationsEngineering(LNT)MunichUniversityofTechnology(TUM)

    D-80290Munich,Germany

    THOMASWIEGAND

    ImageProcessingDepartmentHeinrichHertzInstitute(HHI)

    D-10587Berlin,Germany

  • 8/2/2019 Video Transmission in 3G Telephony

    2/6

    thecell,activeusersmustshareresourceswithnewusers

    andifusersexitthecell,theresourcescanbere-allocated

    tothe remaining users.Thewell-designed3G airinter-

    facesallowdatarateswitchinginaveryflexiblewayby

    assigningappropriatescramblingorchannelcodingrates.

    This results in aneedforvideocodecs tobecapableof(forexample)doublingorhalvingvideodatarateevery5-

    20seconds.Hence,dueto thetime-varyingnatureofthe

    mobilechannel,thevideoapplicationmustbecapableofreactingtovariablebit-rate(VBR)channelsaswellasto

    residualpacketlosses.Finally,theprioritizationandqual-ityofservicedesignformobilelinksisanongoingstan-

    dardizationandresearchactivity.Systemssupportingpri-

    oritizedtransmissionshowimprovedperformanceifvideo

    standardsallowgeneratingdatawithdifferentpriorities.

    B. CommonTestConditionsfor3GMobileVideo

    IntheJVTstandardizationprocesstheimportanceofmo-

    bilevideotransmissionhasbeenrecognizedbyadopting

    appropriatecommontestconditionsfor3Gmobiletrans-missionforcircuitswitchedconversationalservicesbased

    on H.324M [6] and for packet switched conversational

    andstreamingservices[7].Thesetestconditionsallowforselectingappropriatecodingfeatures,totestandevaluate

    errorresiliencefeaturesandtoproducemeaningfulanchor

    streams. In the following we will focus on packet-switchedapplicationsandthecorrespondingcommontest

    conditions,asthisseemstobemoreimportantinnowa-

    days IP-based world. Additionally, the packetization

    scheme as well as the performance for H.324M based

    videotransmissioniscomparabletoIP-basedschemes.

    The common test conditions define video test se-

    quenceswiththeappropriatetemporalandspatialresolu-tion. Additionally, a simplified offline 3GPP/3GPP2simulationsoftware[8]isavailableincombinationwith

    appropriateparametersettings.Forsimulatingradiochan-

    nelconditions,bit-errorpatternsareusedthatwerecap-turedindifferentrealoremulatedmobileradiochannels.

    The bit-error patterns aremeasured above the physical

    layerandbelowtheRLC/RLPlayer,suchthatinpractice

    theyactasthephysicallayersimulation.

    NALpacket

    IP

    Framing,ROHC

    Linklayer

    Physicallayer

    UDP RTP NALpacket RTP/UDP/IP

    RoHCPPP

    frameRLP frameRLP

    PhysicalframeLTU

    frameRLP

    PhysicalframeLTU CRCCRC

    .

    Figure1Packetizationthrough3GPP2protocolstack

    AccordingtoFigure1thesoftwaresimulatorassumesa

    JVT Network Adaptation Layer Packet (NALP) to be

    encapsulated in an IP/UDP/RTP packet at the input.

    NALP usually contain a single slice packet (SSP)ora

    data partition. Formoredetailswe referto [11].In the

    following we briefly examine the user plane protocolstackfor3GPP2CDMA-2000.The3GPPUMTSstackis

    verysimilar.AfterRobustHeaderCompression(RoHC),

    the IP/UDP/RTP packet is encapsulated into one

    PDCP/PPPpacketthatwillbecomeanRLC-SDU.Video

    packetsare in generalof varying length,soRLC-SDUs

    willbeofvaryinglengthaswell.InthecasethatanRLC-

    SDUislargerthananRLC-PDU,theSDUissegmentedintoseveralPDUs.The flowofvariable sizeRLC-SDUs

    iscontinuoustoavoidpaddingbits.RLC-SDUswithone

    ormoreRLC-PDUs thatcontainpart of theRLC-SDUhave not been received correctly are discarded. The

    RLC/RLP layer can perform re-transmissions. The re-transmissionschememaybesetupwithdifferentlevels

    of persistency. The common test conditions specify 12

    anchors with different video sequences, radio bit error

    patterns,transmissionbit-ratesandretransmissionmodes.

    III. JVTCODINGSTANDARD

    A. OverviewCodingAlgorithm

    Although thedesignof theJVTcodec basically fol-lowsthedesignofpriorvideocodingstandardsasMPEG-

    2,H.263, andMPEG-4, it containsmany new features

    that enable it to achieve a significant improvement in

    termsofcompressionefficiency.Wewillbrieflyhighlight

    those.Formoredetailswe refer to [1] and [3]. In JVT

    Coding, blocks of 4x4 samples are used for transform

    coding,andthusaMB(MB)consistsof16luminanceand

    4blocksforeachchrominancecomponent.Conventional

    picture types knownas I-and P-pictures aresupported.

    Furthermore, JVTCodingsupportsmulti-framemotion-

    compensated prediction. That is, more than one prior

    coded picture can be used as reference for the motion

    compensation.Encoderanddecoderhavetostorealreadycoded pictures in a multi-frame buffer. A generalized

    frame-buffering concepthasbeen adoptedallowingmo-

    tion-compensated prediction not just from previous

    frames but also from future frames.For that, a flexible

    andefficientsignalingmethodhasbeenadopted.Inaddi-

    tion, JVT Coding permits so-called multi-hypothesis

    (MH)pictures,whichsimilartoB-Picturesallowtwopre-

    dictionsignalsperblockbutreferencemorethanonepic-

    ture.Therefore, thesimpleB-picture functionalityis in-

    cludedwithMH-pictures.

    AMBcanalwaysbecodedinoneofseveralINTRA-

    modes. There are two classesof INTRAcodingmodes,

    onewhichbasicallyallowstocodeflatregionswithlow

    frequencycomponentsandonewhichallowstocodede-

    tailsinaveryefficientwayutilizingpredictioninthespa-

    tialdomain by referring neighboring samplesofalready

    coded blocks. In addition to the INTRA-modes, various

    efficient INTER-modesare specified in JVTCoding.In

    addition to theSKIP-mode that means justcopyingthe

    contentfromthesamepositionfromthepreviouspicture,

  • 8/2/2019 Video Transmission in 3G Telephony

    3/6

    seven motion-compensated coding modes are available

    for MBs in P-pictures. Eachmotion-compensatedmode

    corresponds toa specific partition of theMB intofixed

    sizeblocksusedformotiondescription.Currently,blocks

    withsizesof16x16,16x8,8x16,8x8,8x4,4x8,and4x4

    samples aresupported by thesyntax,and thusupto16motionvectorsmaybetransmittedforaMB.

    TheJVTCodingsyntaxsupportsquarter-andeighth-

    sampleaccuratemotioncompensation.Themotionvector

    components aredifferentiallycodedusingeithermedian

    or directional prediction from neighboring blocks. The

    chosen prediction depends on the block shape and the

    positioninsidetheMB.JVTCodingisbasicallysimilarto

    otherpriorcodingstandardsinthatitutilizestransform

    codingof thepredictionerror signal.However, in JVT

    Coding the transformationis applied to4x4blocksand,

    insteadoftheDCT,JVTCodingusesaseparableinteger

    transform with basically the same properties as a 4x4

    DCT.Sincethe inversetransformisdefinedbyexactin-teger operations, inverse-transform mismatches are

    avoided.AppropriatetransformsareusedtothefourDC-

    coefficientsof eachchrominancecomponent(2x2 trans-

    form)andtheINTRA16x16-mode(repeated4x4).

    For the quantization of transform coefficients, JVT

    Coding uses scalarquantization. The quantizersarear-

    rangedinawaythatthereisanincreaseofapproximately

    12.5%fromonequantizationparameter(QP)tothenext.

    Thequantizedtransformcoefficientsarescannedinazig-

    zag fashion and converted into codingsymbolsbyrun-

    lengthcoding(RLC).AllsyntaxelementsofaMBinclud-

    ingthecodingsymbolsobtainedafterRLCareconveyed

    byentropycodingmethods.

    JVTCodingsupports twomethodofentropycoding.

    The first one called Universal VariableLengthCoding

    (UVLC) uses one single infinite-extend codeword set.

    InsteadofdesigningadifferentVLCtableforeachsyntax

    element, onlythemapping to thesingleUVLC tableis

    customizedaccordingtothedatastatistics.Theefficiency

    ofentropycodingisimprovedifContext-AdaptiveBinary

    ArithmeticCoding(CABAC)isusedthatallowstheas-

    signmentofnon-integernumbersof bits toeachsymbol

    ofanalphabet.Additionally,theusageofadaptivecodes

    permits theadjustmenttonon-stationarysymbolstatistic

    andcontextmodelingallowsforexploitingstatisticalde-pendencies between symbols. For removing block-edge

    artifacts, the JVT Codingdesign includes a deblocking

    filter.TheJVTCodingblockedgefilterisappliedinside

    themotionpredictionloop.Thefilteringstrengthisadap-

    tivelycontrolledbythevaluesofseveralsyntaxelements.

    B. ErrorResilienceFeatures

    Forenhancederrorresilience,thetestmodelallowsinter-

    ruptingspatial,temporalandsyntacticalpredictivecoding

    onaMBbasis.Theprinciplesofeachoftheadoptedfea-

    turesarereasonablywellknownfrompriorvideocoding

    work, particularly from the H.263+, H.263++, andMPEG-4projects.However,thesefeaturesaretakenabit

    furtherintheJVTCodingdesign.

    Temporal resynchronization within a JVTvideobit-

    streamcanbeaccomplishedbyuseofintrapicturerefresh

    (stopping allprediction of data fromone picture toan-

    other),whereasspatialresynchronizationissupportedbyslicestructuredcoding(providingspatially-distinctresyn-

    chronizationpointswithinthevideodataforasinglepic-

    ture).Inaddition,theusageofintraMBrefreshandmul-

    tiple reference frames allows the encoder to introducewell-selectedintraupdatesreferenceframeselection.Ad-

    ditionally,thepacketlengthcanbeadaptedbyappropriate

    groupingofMBs.Fastrateadaptationcanbeaccomplishedbyswitching

    thequantizationfidelityonaMBbasissuchthatareal-

    time encoder can react immediately tovaryingbit rate.

    For streaming of pre-coded sequences, well-designed

    buffering can deal reasonablywellwith varyingbit-rateconditions.Still, bufferoverflows inVBRenvironments

    may not be completely avoidable. For this reason JVT

    Codingdefinesnewpicturetypes,SP-frame[10]andSI-

    frames, toallowswitchingbetweenversionsofa streamwithoutintroducingtheefficiencylossassociatedwithan

    I-frame. Additionally, a syntax-based data partitioning

    schemewithatmost3partitionsperslicewasintroduced

    in[11]allowinglessimportantinformationtobedropped

    intheeventofabufferoverflowortobeusedinconjunc-tionwithnetworkprioritizationorunequalerrorprotec-

    tiontosupportqualityofserviceconceptsinnetworks.

    IV. TESTMODELEXTENSIONS

    A. Overview

    In theprevious section, thecoding and error resilience

    toolsofJVTCodinghavebeenpresented.Ingeneral,thestandarddefinesonlytheappropriatesyntaxforeachin-

    cluded feature. The selection ofappropriate options for

    each application isupto theimplementer.However,in

    ordertojudgeadoptedandproposednormativefeatures,

    appropriate non-normative mechanisms in the encoderanddecodertestmodelarealsoincluded.TheJVTCod-

    ingprojecthascarefullyadopteddifferentnon-normative

    features in the testmodel software to improveperform-

    anceintermsofcodingefficiencyanderrorresilience.

    Theapproachofrate-distortionoptimizedencodingis

    used in theencoder testmodelto selectthe appropriate

    codingoptions.Thismethodiswellknowntoimprovethecodingefficiencysignificantlyincaseofencodersprovid-

  • 8/2/2019 Video Transmission in 3G Telephony

    4/6

    ing many coding options [12]. In JVT Coding, rate-

    distortion optimized coding is used for selecting MB

    modes,referenceframesandmotionvectors.Inaddition,

    ifwehaveknowledgeofcertainchannelcharacteristics,

    e.g.,apacketlossorbiterrorratemodel,thiscanalsobe

    usedintheencoderaswillbediscussedinmoredetailsinthesubsectionBandC.Althoughthesemethodsingen-

    eral require some knowledge on the expected channel

    conditions,theyarerobustinawaythatonlyaroughes-timateonthechannelcharacteristicsissufficient.

    Moreover,thedecoderincludesinadditiontoasimplepreviousframeerrorconcealment(EC)andanadvanced

    ECschemewhichusescorrectlyreceivedinformationto

    estimatethelostinformationinarecursiveway[13].For

    I-pictures a weighted pixel value averaging is applied

    whereas forP-picturesboundary-matching-basedmotion

    vectorrecoveryisutilized.

    B. OptimizedMBModeandReferenceFrameSelection

    ThenumberofintracodedMBsperpicturecanbedeter-minedattheencoderbyvariousmeans.Fortransmission

    in error-free environments, typically the algorithm for

    rate-distortion optimization might sometimes select anintra MB due tocompressionefficiency. In error-prone

    environments,itmaybedesirabletoincreasethenumber

    ofintraMBstolimitthetemporalpropagationofanyer-roroccurred.Thecurrentencodertestmodelonlyallows

    updating an entire row of MBs periodically. However,

    intra codedMBs in generalshowbadcoding efficiency

    andtheiruseshouldthereforebelimited.

    Error-pronetransmissionshowsrandomlossesmaking

    thedecodingresultarandomvariableaswell.Therefore,

    theencodingdecisionsareoptimizedwithregardstotheexpectedvaluesofthedecodingrandomvariable.Hence,a rate-constrainedMBmode and referenceframe selec-

    tionbasedon theexpectedmeansquareerrordistortion

    andtherateisutilized.ThisschemeselectsintraMBup-dates very carefully by trading of distortion versus rate

    givenaprobabilityofchannelerrors.Theexpecteddecod-

    ingdistortionisestimatedbycomputingthesampleaver-

    ageofthedecodingrandomvariableviarunning Nchan-

    nel-decoderpairsintheencoderinparalleltosimulatethe

    statistics of the channel and its impacton thedecoded

    video.Thisprovidesan estimateoftheexpecteddecoder

    distortionintheencoder.Inourimplementation,thesta-

    tisticalprocessofloosingapacketismodeledindepend-

    entlyforeachoftheNdecoders.Thepacketlossprocessforeachdecoderisalsoassumedtobei.i.d.,andtheslicelossprobabilitypisassumedtobeknownattheencoder.

    Tobemoreprecise, let usdefine thesetof possible

    combinationsofMBmodesand referenceframesasSMB

    includingtheoptiontocodeaMBinINTRA,i.e.without

    temporalerrorpropagation,orinINTER,i.e.,withtempo-

    ralerrorpropagationbuthighercodingefficiency.Then,

    foreachMBtheMBmodemisselectedaccordingto

    !

    " #

    $

    withDmbeingthedistortioninthecurrentMBwhense-

    lectingMBmodemandRmbeingthecorrespondingrate,i.e.thenumberofbits.ThedistortionDmiscomputedas

    % & '

    ( (

    0

    1 2

    3

    4 5 5 6

    4

    6 5

    8 9 9

    @ A

    B C

    D D

    withfibeingtheoriginalpixelvalueatposition iwithin

    theMBandE E

    G H I Q

    S

    beingthereconstructedpixelvalueat

    positioniforcodingMBmodeminthesimulatedchan-

    nel-decoder pair n. Thedecoder in theencoder applies

    simplepreviousframeECandthereforeservesasanup-

    per bound on the expectedMBdistortion.Accordingto

    [14]theparametershoulddependonthequantization

    parameterqas=5exp(0.1q)(q+5)/(34-q).Obviouslyfor

    largeN theencoderhas a good estimateof theaveragedecoder distortion.However,with increasingN alinear

    increaseofstorageandcomputationalcomplexityinthe

    encoderisobvious.Therefore,thismethodmightnotbe

    practical in real-time encodingprocesses. Less complex

    algorithms with similar performance are known [15].

    However, the applications of these algorithms are not

    straightforward due to sub-pelmotion estimation, loop-

    filtering and intraMBprediction inJVTCoding.Addi-

    tionally, for thepurposeof standardization, this simple

    solutionprovidesflexibilityasnewfeaturesinthedecoder

    arejustcopiestotheencoder.Inaddition,anyothererror

    resiliencetoolscanbetestedandanestimationoftheex-

    pecteddecoderdistortioncanbeobtainedintheencodereasily.

    C. PacketLengthSelection

    Anothercriticalparameteristheselectionofthepacket

    length. the JVT Coding standard allows to group any

    numberofMBsintoonesliceorNALpacket.Nospatial

    predictionoversliceboundariesisallowed.Therefore,a

    newsliceallowsresynchronizationwithinoneframe.Inmobileenvironments,theprobabilitythatshorterpackets

    arehitby abit error istypically smaller thanforlarger

    packets.Inaddition,shorterpacketsprovidemoreresyn-

    chronizationpossibilitiesand, therefore,arefavorablein

    termsoferrorresilience.Butsmallerpacketsalsoresultin

    efficiency loss due to the restrictedspatialpredictioninsmallerslicesandtheintroducedsliceheaderandnetwork

    overheadforeachpacketandduetotheinterruptionofthe

    prediction at the source coder at the packet boundary.

    Thoughtheheadersizescouldbereducedbytheintroduc-

    tionof parameter sets [11]and RoHC, they arestillnotnegligible.Hence,acarefulselectionofthepacketlength

    adaptedtochannelandvideoconditionsisvital.

  • 8/2/2019 Video Transmission in 3G Telephony

    5/6

    TheJVTCodingtestmodelencoderallowschoosing

    theslicesizeindifferentmodes.ThenumberofMBsin

    eachslicecanbespecified.Therefore,thepacketsusually

    differinlength.Especiallypacketscontainingintrainfor-

    mationor highmotionareasresult in bigger size,and,

    therefore,aremoresusceptibletobiterrors.Therefore,ina different mode themaximum numberof bytes in one

    packet canbe specified. This assures thatpackets have

    similarlengthand,therefore,arealmostequallysuscepti-ble tobit errorsand resultinglosses.Thepacket length

    canbeadaptedsothatthelossprobabilityisbelowacer-tainthresholdifthechannelconditionsareknownatthe

    encoder.Obviously,anindividualpacketlengthselection

    adaptedtovideocontentanderrorcharacteristicssimilar

    tothepreviouslypresentedadaptiveintraupdatecanhelp

    totradeoffoverheadversuscompressionefficiencyinan

    optimizedway.Thistopicispartofongoingresearch.

    V. SELECTEDSIMULATIONRESULTS

    A. SimulationParameters

    Wewillbrieflypresentsimulationsforselectedparame-

    ters.AnIPbasedconversationalserviceatamobilespeed

    of3km/hat64kbit/sissimulated.Weassumeinteractiveserviceswithasmallend-to-enddelayandtherefore,no

    radioretransmissionsareapplied.Thebearerbiterrorrate

    is at about510-4. TheQCIF videotest sequences Fore-

    man (10s, 300 frames) and Hall Monitor (10s, 300

    frames)aretransmittedataframerateof7.5fpsand15

    fps, respectively [7]. Since no rate control is currently

    present in the JVT video encoder, a fixed quantization

    parameter is selected so that the total video bit-rate

    includingthepacketizationoverheaddoesnotexceed64kbit/s.Foreachsequence,50decodingrunsareperformed

    where each run starts at a different predefined starting

    positionin thebiterror file.Thepacketlossprobability

    obviouslydependsonthepacketlength.Anevaluationofthebiterrorpatternfileshowsforexamplethattheloss

    probabilityofa packet of length200bytesis about2%,

    whereas for a packet of length 500bytes itincreasesto

    about5%.

    B. SimulationResults

    Different experiments have been carried out. We only

    reportonaselectedsubsetandshowthebenefitsofdif-

    ferent JVTCoding errorresilientmodesand testmodel

    extensions. The entropy coding method applied in allcasesisthesimplerUVLC.NotethatapplyingCABACwould result in even better results for all experiments.Extensivesimulation resultsaswell asselected decodedvideosequencesareavailable

    1.Forbothvideosequences,

    wereportforeachinvestigatedcasetheluminancePSNR

    averagedoverall framesand all runs (av)foradvanced

    1Availableathttp://www.ei.tum.de/~stockhammer

    EC(AEC).AdditionallyforForeman,theaverageoverall

    framesfortheworst-case(wc)runarereported.Further-

    more,resultsforpreviousframeEC(PFEC)aregivenas

    well,sinceforthissequence,the twoconcealmentmeth-

    odsprovidequitedifferentresults,whiletheconcealment

    itselfdependsonthenumberofMBswithinaslice.Theresults for different encoder and decoder settings are

    showninTable1.

    Table1ResultsinPSNR(indB)forForeman(FM)and

    HallMonitor(HM)fordifferentencoder/decodersettings

    Exp QP PFECFM AECFM HM

    FM HM av

    PSNR

    wc

    PSNR

    av

    PSNR

    wc

    PSNR

    av

    PSNR

    1 17 13 26.441 15.607 26.441 15.607 32.683

    2 19 19 29.380 24.781 29.380 24.781 34.110

    3 23 16 30.094 28.637 30.094 28.637 35.398

    4 20 18 29.641 22.436 30.519 24.025 33.158

    5 22 25 30.436 28.911 30.701 29.463 30.0366 23 19 30.131 26.168 30.377 28.677 33.887

    7 21 16 30.974 29.215 31.169 30.046 35.628

    8 22 16 30.719 29.152 30.762 29.890 35.625

    InExperiments1,2and3anentireframeistransmit-

    ted in onepacket. Therefore,bothECschemesperform

    identically as theAEConlyexploitsspatial correlationswithin one frame. In experiment 1 no error resilience

    toolshavebeenapplied.TheresultsfortheaveragePSNR

    areacceptableasfortheForemansequencetheR-Dopti-

    mized mode selection selects the intraMB mode quite

    frequently.However,theworst-caseperformanceisvery

    poor indicating that without error resilience encoding

    methods,averybaddecodingqualitymightoccurocca-sionally.Inexperiment2,theintroductionofregularintraupdates (periodically 1 row of MBs is updated every

    frame) providesa significant improvement. Even better

    results canbeobtainedby adaptiveintraupdatesas ap-pliedinexperiment3.Especiallytheworst-casePSNRis

    increasedsignificantlyand,therefore,thevarianceofthe

    receiverquality is reduced.A frame loss rateof 10%is

    assumedcausingalargenumberintraMBupdates.Please

    notethatalsothequantizationparameterisincreaseddue

    tothelowercodingefficiencyofINTRAcoding.Itcanbe

    seenthattheadaptivecodingschemealsoadaptstheQP

    appropriately.FortheForemansequencemanymoreintra

    MBsareusedfortheadaptiveintraupdatecomparedto

    theregularintraupdate.TheQPmatchingtherequiredbitrateismuchhigherforexperiment3comparedtoexperi-ment 2.For theHallMonitor sequence this isdifferent

    thoughthechannelstatisticsareequivalent.Therefore,it

    is obvious that the redundancy necessary to copewith

    packetlossisadaptednotonlytothechannelstatisticsbut

    alsotothevideocontent.Thisshowsthevalidityandim-

    portanceoftherate-distortionapproach.

  • 8/2/2019 Video Transmission in 3G Telephony

    6/6

    Inexperiments4-8sliceshavebeenintroducedtoob-

    tain shorter packets. In these experiments,typically the

    AEC provides gains as spatial correlations within one

    frame canbeexploitedat thedecoder.Experiment4is

    equivalenttoexperiment1exceptthateachrowofMBsis

    transmittedinaseparateslice.Thisallowsfrequentresyn-chronization,butalso reducescodingefficiency.These-

    lectedQPinthiscaseis20fortheForemansequence,18

    for theHall Monitor sequence. The performance com-paredtoexperiment1isslightlyincreased,especiallyfor

    theAEC. Introducing regular intra updates as done inexperiment5doesimprovethequalitysignificantlyfor,

    butnot for theHallMonitorsequence.FortheForeman

    sequencetheamountofintraupdatesforexperiment4fits

    quitewellthevideocontentandchannelstatistics.

    Inexperiments6,7and8thecombinationofadaptive

    intra updatesand the slicefeature is investigated. Inall

    cases it canbe seen thattheAEC canimprove thede-

    coded quality, especially in theworst scenarios. In ex-

    periment 6 the slice loss probability at the encoder is

    assumedtobe3%andtheslicelengthwasthesameasinexperiment4and5,i.e.onerowofMBsinoneslice.Thegains in this case are significant compared to allothernon-adaptivecases,forPFECaswellasforAEC.Againit

    canbeobservedthatfortheForemansequencethesetting

    in experiment5 iswellmatchedand theadaptive intra

    updateisslightlyworseforexperiment6.Thisis,asthe

    channel statistics cannot be modeled accurately in the

    encoderwithanindependentpacketlossmodel.However,

    fortheHallMonitorsequencethisexperimentshowssig-

    nificant gainscomparedto experiment5. Therefore,the

    robustness of adaptive intra updates is obvious. In ex-

    periment7,theslicelengthissetto33MBssowehave3packetspervideoframe.Alossprobabilityof5%isas-sumedattheencoder.Theresultsofthisexperimentout-

    perform all other experiments as thetradeoff of packet

    overhead,intraupdatesandlossrateiswellmatched.The

    AECimprovestheresultsslightlyasspatialcorrelations

    canbeexploitedifnotallpacketsofoneframearelost.

    Finally,inexperiment8,themaximumpacketlength

    islimitedto256bytesandthelossprobabilityassumedat

    theencoderis5%.Theperformanceisverysimilartothe

    resultsinexperiment7.However,ingeneraltheapproachlimitingthepacketlengthratherfixingthenumberofMB

    canprovidebetterresult.

    VI. CONCLUSIONSANDOUTLOOK

    The JVT Codingprojectpromises somesignificant ad-

    vances in the state-of-the-artof standardizedvideocod-

    ing,includingkeyaspectsdesignedwithmobileapplica-tionsin mind. In addition toexcellent codingefficiency

    with halving thebit-ratecompared to all existing stan-

    dards, the JVT Coding project also takes into account

    networkadaptationintheinherentdesign.Thisincludes

    the definitionof appropriatemobile testconditions, the

    integrationofnetworkrelatedanderror-resilientfeatures

    inthestandardand,finally,theextensionofthetestmodel

    software tofullyexploit theintegratedfeatures.This al-

    lowstoprovideappropriatejudgmentofadoptedandpro-

    posedfeatures andto finally cometoa standard,whichwill help toimprovethe quality of lowbit-ratevideoin

    3Gmobileenvironmentssignificantly.Furtherworkwill

    beconductedwithintheJVTtoimprovecodingefficiencyas well as network and application friendliness and to

    provideastandardsuitablefordifferentapplicationsandnetworkswithspecialfocusonmobilevideoapplications.

    REFERENCES

    [1] T.Wiegand, WorkingDraftNumber1, Revision0 (WD-1),Joint

    Video Team (JVT) of ISO/IEC MPEG and ITU-T VCEG, JVT-

    A003,January2002.

    [2] ITU-T, Video Coding for Low Bit-Rate Communication, ITU-T

    Recommendation H.263, Version 1: November 1995, Version 2:

    January1998,Version3:Nov.2000.[3] G. Sullivan, T. Wiegand, and T. Stockhammer, Using the Draft

    H.26L Video Coding Standard for Mobile Applications, in Proc.

    ICIP2001,Thessaloniki,Greece,October2001.

    [4] ISO/IEC JTC1, Generic Coding of Audiovisual Objects Part 2:

    Visual (MPEG-4 Visual), ISO/IEC 14496-2, Version 1: January

    1999,Version2:January2000;Version3:January2001.

    [5] S. Wenger, M. Hannuksela, and T. Stockhammer, Identified

    H.26L Applications, ITU-T SG 16, Doc. VCEG-L34, Eibsee,

    Germany,Jan.2001.

    [6] ITU-T Q15-I-60, Common Conditions for Video Performance

    Evaluation in H.324/M error-prone systems, VCEG (SG16/Q15),

    NinthMeeting,Redbank,NJ,October1999.

    [7] G.Roth,R. Sjberg, G.Liebl,T. Stockhammer, V. Varsa, andM.

    Karczewicz, Common Test Conditions for RTP/IP over

    3GPP/3GPP2, ITU-TSG16 Doc. VCEG-M77, Austin,TX, USA,

    Apr.2001.[8] G.Roth,R. Sjberg, G.Liebl,T. Stockhammer, V. Varsa, andM.

    Karczewicz, Common Test Conditions for RTP/IP over

    3GPP/3GPP2 Amendments and Software, ITU-T SG16 Doc.

    VCEG-M37,SantaBarbara,CA,USA,Sept..2001.

    [9] D. Marpe, G. Blttermann, G. Heising and T. Wiegand, Video

    Compression using Context-Based Arithmetic Coding, in Proc.

    ICIP2001,Thessaloniki,Greece,October2001.

    [10]R.Kurceren and M.Karczewicz,AProposalforSP-frames,ITU-

    TSG16Doc.VCEG-L27,Eibsee,Germany,Jan.2001.

    [11]S.Wenger andT.Stockhammer,H.26L over IPand H.324Frame-

    work, ITU-TVCEG-N52,VCEG (SG16/Q6), FourteenthMeeting,

    SantaBarbara,CA,September2001.

    [12]G.J. Sullivan and T. Wiegand, Rate-Distortion Optimization for

    Video Compression, IEEE Signal Processing Magazine, vol. 15,

    no.6,pp.74-90,Nov.1998.

    [13]V. Varsa, M. Hannuksela, and Y. Wang, Non-normative errorconcealment algorithms, ITU-T VCEG-N62, VCEG (SG16/Q6),

    FourteenthMeeting,SantaBarbara,CA,September2001.

    [14]T. Wiegand, and B. Girod, Lagrangian Multiplier Selection in

    Hybrid Video Coder Control, in Proc. ICIP 2001, Thessaloniki,

    Greece,October2001.

    [15]R.Zhang,S.L.Regunathan,andK.Rose,VideoCodingwith

    OptimalInter/Intra-ModeSwitchingforPacketLossResilience,in

    IEEEJSAC,vol.18,no.6,pp.966-976.