VP8 Data Format and Decoding Guide.docx

VP8 Data Format and Decoding Guide.docx

《VP8 Data Format and Decoding Guide.docx》由会员分享，可在线阅读，更多相关《VP8 Data Format and Decoding Guide.docx（100页珍藏版）》请在冰豆网上搜索。

VP8DataFormatandDecodingGuide

VP8DataFormatandDecodingGuide

WebMProject

Google,Inc.

3CorporateDrive,Suite100

CliftonPark,NY12065USA

http:

//www.webmproject.org/

Revised:

May2010

License

Copyright

Chapter1:

Introduction

Chapter2:

FormatOverview

Chapter3:

CompressedFrameTypes

Chapter4:

OverviewofCompressedDataFormat

Chapter5:

OverviewoftheDecodingProcess

Chapter6:

DescriptionofAlgorithms

Chapter7:

BooleanEntropyDecoder

7.1UnderlyingTheoryofCoding

7.2PracticalAlgorithmDescription

7.3ActualImplementation

Chapter8:

CompressedDataComponents

8.1TreeCodingImplementation

8.2TreeCodingExample

Chapter9:

FrameHeader

9.1UncompressedDataChunk

9.2ColorSpaceandPixelType（KeyFrames-only）

9.3Segment-basedAdjustments

9.4LoopFilterTypeandLevels

9.5TokenPartitionandPartitionDataOffsets

9.6DequantizationIndices

9.7RefreshGoldenFrameandAltRefFrame

9.8RefreshLastFrameBuffer

9.9DCTCoefficientProbabilityUpdate

9.10RemainingFrameHeaderData（non-KeyFrame）

9.11RemainingFrameHeaderData（KeyFrame）

Chapter10:

Segment-basedFeatureAdjustments

Chapter11:

KeyFrameMacroblockPredictionRecords

11.1mb_skip_coeff

11.2LumaModes

11.3SubblockModeContexts

11.4ChromaModes

11.5SubblockModeProbabilityTable

Chapter12:

IntraframePrediction

12.1mb_skip_coeff

12.2ChromaPrediction

12.3LumaPrediction

Chapter13:

DCTCoefficientDecoding

13.1MBWithoutnon-ZeroCoefficientValues

13.2CodingofIndividualCoefficientValues

13.3TokenProbabilities

13.4TokenProbabilityUpdates

13.5DefaultTokenProbabilityTable

Chapter14:

DCTandWHTInversionandMacroblockReconstruction

14.1Dequantization

14.2InverseTransforms

14.3ImplementationoftheWHTInversion

14.4ImplementationoftheDCTInversion

14.5SummationofPredictorandResidue

Chapter15:

LoopFilter

15.1FilterGeometryandOverallProcedure

15.2SimpleFilter

15.3NormalFilter

15.4CalculationofControlParameters

Chapter16:

InterframeMacroblockPredictionRecords

16.1Intra-PredictedMacroblocks

16.2Inter-PredictedMacroblocks

16.3ModeandMotionVectorContexts

16.4SplitPrediction

Chapter17:

MotionVectorDecoding

17.1CodingofEachComponent

17.2ProbabilityUpdates

Chapter18:

InterframePrediction

18.1BoundsonandAdjustmentofMotionVectors

18.2PredictionSubblocks

18.3Sub-pixelInterpolation

18.4FilterProperties

Chapter19:

References

RevisionHistory

READMEforVP8BitstreamSpecification

History

Format

MathML

License

GoogleherebygrantstoYouaperpetual,worldwide,non-exclusive,no-charge,royalty-free,irrevocable（exceptasstatedinthissection）patentlicensetomake,havemade,use,offertosell,sell,import,andotherwiseimplementationsofthisspecificationwheresuchlicenseappliesonlytothosepatentclaims,bothcurrentlyownedbyGoogleandacquiredinthefuture,licensablebyGooglethatarenecessarilyinfringedbyimplementationofthisspecification.IfYouoryouragentorexclusivelicenseeinstituteororderoragreetotheinstitutionofpatentlitigationagainstanyentity（includingacross-claimorcounterclaiminalawsuit）allegingthatanyimplementationofthisspecificationconstitutesdirectorcontributorypatentinfringement,orinducementofpatentinfringement,thenanyrightsgrantedtoYouundertheLicenseforthisspecificationshallterminateasofthedatesuchlitigationisfiled.

Copyright

ThisspecificationismadeavailableunderaCreativeCommonsAttribution3.0License.

Chapter1:

Introduction

ThisdocumentdescribestheVP8compressedvideodataformatcreatedbyGoogleOn2,togetherwithadiscussionofthedecodingprocedureforthisformat.ItisintendedtobeusedinconjunctionwithandasaguidetothereferencedecoderprovidedbyGoogleOn2.Ifthereareanyconflictsbetweenthisdocumentandthereferencesourcecode,thereferencesourcecodeshouldbeconsideredcorrect.Thebitstreamisdefinedbythereferencesourcecodeandnotthisdocument.

Likemanymodernvideocompressionschemes,VP8isbasedondecompositionofframesintosquaresubblocksofpixels,predictionofsuchsubblocksusingpreviouslyconstructedblocks,andadjustmentofsuchpredictions（aswellassynthesisofunpredictedblocks）usingadiscretecosinetransform（hereafterabbreviatedasDCT）.Inonespecialcase,however,VP8usesa“Walsh-Hadamard”（hereafterabbreviatedasWHT）transforminsteadofaDCT.

Roughlyspeaking,suchsystemsreducedataratebyexploitingthetemporalandspatialcoherenceofmostvideosignals.Itismoreefficienttospecifythelocationofavisuallysimilarportionofapriorframethanitistospecifypixelvalues.ThefrequencysegregationprovidedbytheDCTandWHTfacilitatetheexploitationofbothspatialcoherenceintheoriginalsignalandthetoleranceofthehumanvisualsystemtomoderatelossesoffidelityinthereconstitutedsignal.

VP8augmentsthesebasicconceptswith,amongotherthings,sophisticatedusageofcontextualprobabilities.Theresultisasignificantreductionindatarateatagivenquality.

Unlikesomesimilarschemes（theolderMPEGformats,forexample）,VP8specifiesexactvaluesforreconstructedpixels.Specifically,thespecificationfortheDCTandWHTportionsofthereconstructiondoesnotallowforany“drift”causedbytruncationoffractions.Rather,thealgorithmisspecifiedusingfixedprecisionintegeroperationsexclusively.Thisgreatlyfacilitatestheverificationofthecorrectnessofadecoderimplementationaswellasavoidingdifficult-to-predictvisualincongruitiesbetweensuchimplementations.

Itshouldberemarkedthat,inacompletevideoplaybacksystem,thedisplayedframesmayormaynotbeidenticaltothereconstructedframes.Manysystemsapplyafinalleveloffiltering（commonlyreferredtoaspostprocessing）tothereconstructedframespriortoviewing.Suchpostprocessinghasnoeffectonthedecodingandreconstructionofsubsequentframes（whicharepredictedusingthecompletely-specifiedreconstructedframes）andisbeyondthescopeofthisdocument.Inpractice,thenatureandextentofthissortofpostprocessingisdependentonboththetasteoftheuserandonthecomputationalfacilitiesoftheplaybackenvironment.

Chapter2:

FormatOverview

VP8worksexclusivelywithan8-bitYUV4:

2:

0imageformat.Inthisformat,each8-bitpixelinthetwochromaplanes（UandV）correspondspositionallytoa2x2blockof8-bitlumapixelsintheYplane;coordinatesoftheupperleftcorneroftheYblockareofcourseexactlytwicethecoordinatesofthecorrespondingchromapixels.Whenwerefertopixelsorpixeldistanceswithoutspecifyingaplane,weareimplicitlyreferringtotheYplaneortothecompleteimage,bothofwhichhavethesame（full）resolution.

Asisusuallythecase,thepixelsaresimplyalargearrayofbytesstoredinrowsfromtoptobottom,eachrowbeingstoredfromlefttoright.This“lefttoright”then“toptobottom”raster-scanorderisreflectedinthelayoutofthecompresseddataaswell.

ProvisionhasbeenmadeforthesupportoftwodifferentYUVcolorformatsintheVP8bitstreamheader,howeveronlyoneformatissupportedinthefirstreleaseofVP8.

TheYUVformatsdifferintermsoftheirconversiontoandfromRGBcolorspace.ThefirstcorrespondstothetraditionalYUVcolorspacesimilartotheYCrCbcolorspacedefinedinITU-RBT.601.Thesecond（currentlyunsupported）formatcorrespondstoanewYUVcolorspacewhosedigitalconversiontoandfromRGBcanbeimplementedwithoutmultiplicationsanddivides.TheVP8DecodershoulddecodeandpasstheinformationontotheprocessesthatconverttheYUVoutputtoRGBcolorspace.

Occasionally,atverylowdatarates,acompressionsystemmaydecidetoreducetheresolutionoftheinputsignaltofacilitateefficientcompression.TheVP8dataformatsupportsthisviaoptionalupscalingofitsinternalreconstructionbufferpriortooutput（thisiscompletelydistinctfromtheoptionalpostprocessingdiscussedearlier,whichhasnothingtodowithdecodingperse）.Thisupsamplingrestoresthevideoframestotheiroriginalresolution.Inotherwords,thecompression/decompressionsystemcanbeviewedasa“blackbox”,wheretheinputandoutputisalwaysatagivenresolution.Thecompressormightdecideto“cheat”andprocessthesignalatalowerresolution.Inthatcase,thedecompressorneedstheabilitytorestorethesignaltoitsoriginalresolution.

Internally,VP8decomposeseachoutputframeintoanarrayofmacroblocks.AmacroblockisasquarearrayofpixelswhoseYdimensionsare16x16andwhoseUandVdimensionsare8x8.Macroblock-leveldatainacompressedframeoccurs（andmustbeprocessed）inarasterordersimilartothatofthepixelscomprisingtheframe.

Macroblocksarefurtherdecomposedinto4x4subblocks.Everymacroblockhas16Ysubblocks,4Usubblocks,and4Vsubblocks.Anysubblock-leveldata（andprocessingofsuchdata）againoccursinrasterorder,thistimeinrasterorderwithinthecontainingmacroblock.

Asdiscussedinfurtherdetailbelow,datacanbespecifiedatthelevelsofbothmacroblocksandtheirsubblocks.

Pixelsarealwaystreated,ataminimum,atthelevelofsubblocks,whichmaybethoughtofasthe“atoms”oftheVP8algorithm.Inparticular,the2x2chromablockscorrespondingto4x4Ysubblocksarenevertreatedexplicitlyinthedataformatorinthealgorithmspecification.

TheDCTandWHTalwaysoperateata4x4resolution.TheDCTisusedforthe16Y,4Uand4Vsubblocks.

TheWHTisused（withsomebutnotallpredictionmodes）toencodea4x4arraycomprisingtheaverageintensitiesofthe16Ysubblocksofamacroblock.The