università degli studi roma tre – introduzione alla codifica h.264/avc università degli studi...

Università degli studi Roma Tre – Introduzione alla codifica H.264/AVC

Università degli studi Roma Tre

Overview of the H.264AVC video coding standard

Maiorana Emanuele


Agenda

• Introduction

• Project overview & timeline

• Standardization concepts

• Codec technical design

• New Fewtures Prediction De-Blocking Entropy Coding

• Profiles & levels

• Comparisons


Introduction

H.264/AVC is newest video coding standard developed by the ITU-T/ISO/IEC Joint Video Team (JVT), consisting of experts from:

• ITU-T Video Coding Experts Group (VCEG)• ISO/IEC Moving Picture Experts Group (MPEG)

Its design represents a delicate balance between:

• Coding Gain (improved efficiency by a factor of two over MPEG-2)• Implementation complexity• Costs based on state of VLSI (ASICs and Microprocessors) design

technology


Terminology

The following terms are used interchangeably:

• H.26L• The Work of the JVT or “JVT CODEC” • JM2.x, JM3.x, JM4.x• The Thing Beyond H.26L• The “AVC” or Advanced Video CODEC

Proper Terminology going forward:

• MPEG-4 Part 10 (Official MPEG Term) ISO/IEC 14496-10 AVC

• H.264 (Official ITU Term)


History

The digital representation of the TeleVision signals created many services for the content delivery:

• Satellite• Cable TV• Terrestrial Broadcasting• ADSL and Fiber on IP

To optimize this services, there is the need of:

• High Quality of Service (QoS)• Low Bit-Rate• Low Power Consumption

The source coding is responsible for the reduction of the bit-rate.

Example: the complete transmission of the TV signal, as in Recommendation ITU-R BT.601, would require:

720 × 576 + 2 (360 × 576) × 25 × 8= 166 Mbit/s

Conflicting


Video Coding History

Efforts in maximizing coding efficiency while dealing with:

• diversification of network types• characteristic formatting and loss/error robustness requirements.

ITU Standard for VideoTelephony

• H.261, H.263, H.263+

ISO-MPEG Standard

• MPEG-1:medium quality, physical support

• MPEG-2:medium/high quality, physical and transmission support

• MPEG-4:audio video objects


Evolution (1/2)

0

1

2

3

4

5

6

1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005

M

bit

/s

MPEG-2

0

1

2

3

4

5

6

1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005

Mb

it/s

MPEG-2

MPEG-4

H.263

MPEG-2 Introduction

MPEG-4 in Comparison


Evolution (2/2)

H.26L Provides Focus

MPEG-4 “Adopts” H.26L

0

1

2

3

4

5

6

1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005

M

bit

/s

MPEG-2

MPEG-4

H.26L

H.263

0

1

2

3

4

5

6

1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005

Mb

it/s

MPEG-2

MPEG-4

H.26L

H.263


Codec Defects

Blocking Low Original High


Codec Defects

Packet Loss


Video Coding History

• Early 1998: Started as ITU-T Q.6/SG16 (VCEG - Video Coding Experts Group) “H.26L” standardization activity

• August 1999: first draft design

• July 2001: MPEG open call for “AVC” technology: H.26L wins

• December 2001: Formation of the Joint Video Team (JVT) between VCEG and MPEG to finalize H.26L as a joint project similar to MPEG-2/H.262

• July 2002: Final Committee Draft status in MPEG

• March 2003: formal approval submission

• October 2004: final ITU-T and ISO approvation


Versions


JVT Project Technical Objectives

Primary technical objectives:

• Significant improvement in coding efficiency: Average bit rate reduction of 50% compared to any other video standard

• Network-friendly video representation for “conversational” (video telephony) and “non-conversational” (storage, broadcast or streaming) applications

• Error resilient coding• Simple syntax specification and targeting simple and clean solutions

The scope of the standardization is only the central decoder, by:

• imposing restrictions on the bitstream and syntax• defining the decoding process such that every conforming decoder

will produce similar output with an encoded bitstream input

Pre-Processing Encoding

Post-Processing & Error Recovery

Decoding

Source

DestinationScope of the Standard


Applications

The new standard is designed for technical solutions including at least the following application areas

• Broadcast over cable, satellite, Cable Modem, DSL, terrestrial, etc.• Interactive or serial storage on optical and magnetic devices, DVD,

etc.• Conversational services over ISDN, Ethernet, LAN, DSL, wireless and

mobile networks, modems, etc. or mixtures of these.• Video-on-demand or multimedia streaming services over ISDN, Cable

Modem, DSL, LAN, wireless networks, etc.• Multimedia Messaging Services (MMS) over ISDN, DSL, Ethernet,

LAN, wireless and mobile networks, etc.

How to handle this variety of applications and networks?


H.264 Design

To address this need for flexibility and customizability, the H.264/AVC design covers a:

• Video Coding Layer (VCL): representation of the video content (performing all the classic signal processing tasks)

• Network Abstraction Layer (NAL): adaptation of VCL representations in a manner appropriate for conveyance by a variety of transport layers or storage media

Video Coding Layer

H.320

Contr

ol D

ata

Coded Macroblock

Coded Slice/Partition

etc.MPEG-2H323/IPMP4FF

Network Abstraction Layer

Data Partitioning


Features enhancing coding efficiency (1)

Enhancements on picture encoding is enabled through value prediction methods

• Variable block-size motion compensation with small block sizes

• Quarter-sample-accurate motion compensation

• Multiple reference picture motion compensation

• Weighted prediction

• Directional spatial prediction for intra coding

• In-the-loop deblocking filtering


Features enhancing coding efficiency (2)

Enhancements on picture encoding is enabled through high performance tools

• Small block-size transform

• Hierarchical block transform

• Short word-length transform

• Exact-match inverse transform

• Arithmetic entropy coding

• Context-adaptive entropy coding


Features enhancing Robustness

Enhancements on Robustness to data errors/losses and flexibility for operation over a variety of network environments is enabled by new design aspects new

• Parameter set structure

• NAL unit syntax structure

• Flexible slice size

• Flexible macroblock ordering (FMO)

• Redundant pictures

• SP/SI synchronization/switching pictures


Network Abstraction Layer

The Network Abstraction Layer (NAL) is designed in order to provide "network friendliness“, facilitating the ability to map H.264/AVC VCL data to transport layers such as

• RTP/IP for any kind of real-time wire-line and wireless Internet services (conversational and streaming)

• File formats, e.g. ISO MP4 for storage and MMS• H.32X for wireline and wireless conversational services• MPEG-2 systems for broadcasting services, etc.

Some key concepts of the Network Abstraction Layer are:

• NAL Units• Use of NAL Units in:

Byte stream format systems Packet-Transport systems

• Parameter Sets• Access Units


NAL Units

The coded video data is organized into NAL units, each of which is effectively a packet that contains an integer number of bytes

NAL units are classified into

• VCL NAL units: contain the data associated to the video pictures• non-VCL NAL units: contain any associated additional information

Header byte: first byte of each NAL unit; contains an indication of the type of data in the NAL unit, and the remaining bytes contain payload data of the type indicated by the header

Emulation Prevention Bytes: bytes inserted in the payload data to prevent the accidentally generation of a particular pattern of data called a start code prefix

NAL unit stream: series of NAL units generated by an encoder


Use of NAL Units

Bitstream-oriented transport systems (H.320, MPEG-2 systems)

• Delivery of the entire or partial NAL unit stream as an ordered stream of bytes or bits the locations of NAL unit boundaries need to be identifiable

• In the byte stream format, each NAL unit is prefixed by a specific pattern of three bytes called a start code prefix

Packet-oriented transport systems (IP, RTP systems)

• The coded data is carried in packets that are framed by the system transport protocol the boundaries of NAL units within the packets can be established without use of start code prefix patterns

• The NAL units can be carried in data packets without start code prefixes.


Parameters Set

A Parameter Set contains information that is expected to rarely change.

Types of parameter sets:

• sequence parameter sets: relative to a series of consecutive coded video pictures (coded video sequence)

• picture parameter sets: relative to one or more individual pictures.

Parameter sets can be sent

• One time (ahead the VCL NAL Units)• Many time (to provide robustness)• In-band (same VCL NAL Unit Channel)• Out-of-Band (different Channel)

Out-of-band Transmission


Access Units

A set of NAL units in a specified form is referred to as an Access Unit.The decoding of each access unit results in one decoded picture.

It can be composed by:

• access unit delimiter: to aid in locating the start of the access unit.• supplemental enhancement information (SEI): containing data such as picture timing information• primary coded picture: set of VCL NAL units that represent the samples of the video picture.• redundant coded pictures: for use by a decoder in recovering from loss or corruption• end of sequence: if the coded picture is the last picture of a coded video sequence• end of stream: if the coded picture is the last coded picture in the entire NAL unit stream


Video Coding Layer

The VCL design follows the so-called block-based hybrid video coding approach

There is no single coding element in the VCL that provides the majority of the significant improvement in compression efficiency in relation to prior video coding standards.

EntropyCoding

Scaling & Inv. Transform

Motion-Compensation

Control Data

Quant.Transf. coeffs

Motion Data

Intra/Inter

Decoder

MotionEstimation

Transform/Scal./Quant.-

Input Video Signal

Split intoMacroblocks16x16 pixels

Intra-frame Prediction

De-blockingFilter

Output VideoSignal

CoderControl


Video Coding Layer

• The picture is split into blocks.

• Intra coded of the first picture or a random access point

• Inter coding for all remaining pictures or between random access points

• Transmission of the motion data as side information

• Transform of the residual of the prediction (Intra or Inter)

• Quantization of the transform coefficients

• Entropy coding and transmission of the quantized transform coefficients, together with the side information


Pictures, frames, and fields

A coded pictures can represent either an entire frame or a single field

A frame of video can be considered to contain two interleaved fields

• interlaced frame: the two fields of a frame were captured at different time instants

• progressive frame

The coding representation in H.264/AVC is primarily agnostic with respect to this video characteristic


Adaptive frame/field coding operation

In interlaced frames with regions of moving objects, two adjacent rows tend to show a reduced degree of statistical dependency

H.264/AVC design allows any of the following decisions for coding a frame:

• Frame mode: combine the two fields together• Field mode: not combine the two fields together

The choice can be made adaptively for each frame and is referred to as Picture Adaptive Frame/Field (PAFF) coding

Field mode:

• Motion compensation utilizes reference fields• De-blocking filter is not used for horizontal edges of macroblocks

Moving region field mode

Non-moving region frame mode


Sampling

YCbCr color space

H.264/AVC uses a sampling structure called 4:2:0 sampling with 8 bits of precision per sample

The chroma component has one fourth of the number of samples than the luma component (in both the horizontal and vertical dimensions)

Y is called luma, and represents brightness.

Cb and Cr are called chroma, and represent the deviation from gray toward blue and red


Macroblocks and Slices

Fixed-size macroblocks partition with 16x16 samples of the luma component and 8x8 samples of each of the two chroma components.

Slices are a sequence of macroblocks which are processed in the order of a raster scan when not using Flexible Macroblock Ordering (FMO).

A picture is a collection of one or more slices in H.264/AVC.

Indipendency Each slice can be correctly decoded without use of data from other slices.


Flexible Macroblock Ordering

FMO uses the concept of slice groups

A set of macroblocks defined by a macroblock to slice group map, specified in the picture parameter set

A slice is a sequence of macroblocks within the same slice group

Useful for concealment in video conferencing applications


Slice Types

I slice: a slice in which all macroblocks of the slice are coded using intra prediction is coded exploiting only the spatial correlation

P slice: In addition to the coding types of the I slice, some macroblocks of the P slice can also be coded using inter prediction with backward references (I or P slices)

B slice: In addition to the coding types available in a P, some macroblocks of the B slice can also be coded using inter prediction with forward references (I, P or B slices)

The following two coding types for slices are new:

SP slice: a slice that is coded such that efficient switching between different pre-coded pictures becomes possible

SI slice: a slice that allows an exact match of a macroblock in an SP slice for random access and error recovery purposes


Motivation for SP and SI slices

The best-effort nature of today’s networks causes variations of the effective bandwidth available to a user

For Video Streaming, the server should adjusting, on the fly, source encoding parameters

Representation of each sequence using multiple and independent streams

Prior video encoding standardsSwitching is possible only at I-frames.

H.264Identical SP-frames can be obtainedeven when they are predicted usingdifferent reference frames.


Intra-frame Prediction

In all slice-coding types, the following types of intra coding are supported

• Intra_4x4 with chroma prediction: areas of a picture with significant detail

• Intra_16x16 with chroma prediction : very smooth areas of a picture• I_PCM: values of anomalous picture content (accurately

representation)

Intra prediction in H.264/AVC is always conducted in the spatial domain

IDR: picture composed of slice I only

• can be decoded without any reference • no subsequent picture in the stream will require reference to pictures

prior to IDR

Chroma samples: similar prediction technique as for the luma component in Intra_16x16 macroblocks


Intra_4x4 modeM A G HIJKL

a b c de f g hi j k lm n o p

B C D E F

Labelling of prediction samples(4x4)

M A G HIJKL

B C D E F

0 (vertical)

M A G HIJKL

B C D E F

1 (horizontal)

M A G HIJKL

B C D E F

2 (DC)

(A+B+C+D+I+J+K+L)/8

M A G HIJKL

B C D E F

3 (diagonal down-left)


Intra_4x4 modeM A G HIJKL

a b c de f g hi j k lm n o p

B C D E F

Labelling of prediction samples(4x4)

M A G HIJKL

B C D E F

4 (diagonal down-right)

M A GHIJKL

B C D E F

5 (vertical-right)

M A G HIJKL

B C D E F

6 (horizontal-down)

M A G HIJKL

B C D E F

7 (vertical-left)


Intra_4x4 mode

When samples E-H are not available, they are replaced by D

M A G H

IJ

K

L

a b c de f g h

i j k l

m n o p

B C D E F

Labelling of prediction samples( 4x4)

M A G HIJKL

B C D E F

8 (horizontal-up)


Intra_16x16 and I_PCM mode

Intra_16x16• mode 0: vertical• Mode 1: horizontal• mode 2: DC• Mode 3: plane (a linear “plane” function is fitted to the upper and

left-hand samples in areas of smoothly-varying luminance)

I_PCM sends directly the values of the encoded samples, to to precisely represent them


Inter-frame Prediction in P Slices

Partitions with luma block sizes of 16x16, 16x8, 8x16, and 8x8 samples are supported by the syntax.

In case partitions with 8x8 samples are chosen, one additional syntax element for each 8x8 partition is transmitted.

The prediction signal is specified by• a translational motion vector• a picture reference index a maximum of sixteen motion vectors

may be transmitted for a single P macroblock.


Algorithm

The encoder selects the “best” partition size for each part of the frame, to minimize the coded residual and motion vectors.

The macroblock partitions chosen for each area are shown superimposed on the residual frame.

• little change between the frames (residual appears grey) a 16x16 partition is chosen

• detailed motion (residual appears black or white) smaller partitions are more efficient.

Residual (no motion compensation)


Effects (1/2)

Residual (no motion compensation)

Frame Fn Frame Fn-1

Residual (16x16 bock size)


Effects (2/2)

Residual (4x4 bock size; half pixel)

Residual (8x8 bock size) Residual (4x4 bock size)

Residual (4x4 bock size; quarter pixel)


Example (1/2)

Residual Fn – F’n-1 (no motion compensation)

Frame Fn Reconstructed reference Frame F’n-1

16x16 Motion Vector Field


Example (2/2)

Motion compensation reference frame Motion compensation residual frame


Motion Estimation Accuracy

The accuracy of motion compensation is in units of one quarter of the distance between luma samples.

Integer-sample position the prediction signal consists of the corresponding samples of the reference picture

Non integer-sample position the corresponding sample is obtained using interpolation to generate non-integer positions.

The prediction values at half-sample positions are obtained by applying a one-dimensional 6-tap FIR Wiener filter horizontally and vertically


Motion Estimation Accuracy• Half sample positions (aa, bb, b, s, gg, hh and cc, dd, h, m, ee, ff)

are derived by first calculating intermediate values

Ex.: b1 = ( E – 5 F + 20 G + 20 H – 5 I + J )h1 = ( A – 5 C + 20 G + 20 M – 5 R + T )

b = (b1 + 16) >>5h = (h1 + 16) >> 5

• Position jj1 = cc1 – 5 dd1 + 20 h1 + 20 m1 – 5 ee1 + ff1

j = ( j1 + 512) >> 10

• Quarter sample positions (a, c, d, n, f, i, k, q) are derived by averaging with upward rounding of the two nearest samples at integer and half sample positions

Ex.: a = ( G + b + 1 ) >> 1

• Quarter sample positions (e, g, p, r) are derived by averaging with upward rounding of the two nearest samples at half sample positions in the diagonal direction as, for example, by

Ex.: e = ( b + h + 1 ) >> 1


Motion Estimation Accuracy

The prediction values for the chroma component are always obtained by bi-linear interpolation.

For chroma the resolution is halved (4:2:0) therefore the motion compensation precision is down to one-eighth pixel ¼ pixels accuracy

a = round{[(8-dx)·(8-dx)·A]+dx·(8-dy)·B+(8-dx)·dy · C+dx · dy · D ]/64} Ex.: a = round[(30A+10B+18C+6D)/64]


Multi-picture Prediction

Multi-picture motion compensation using previously-encoded pictures as references allows up to 32 reference pictures to be used in some cases

Very significant bit rate reduction for scenes with• rapid repetitive flashing• back-and-forth scene cuts• uncovered background areas


Inter-frame Prediction in B Slices

The concept of B slices is generalized in H.264/AVC

Other pictures can reference pictures containing B slices for motion-compensated prediction

Some macroblocks or blocks may use a weighted average of two distinct motion-compensated prediction values for building the prediction signal

In B slices, four different types of inter-picture prediction are supported:• list 0 (backward)• list 1 (forward)• bi-predictive: weighted average of motion-compensated list 0 and list

1 prediction signals• direct prediction: inferred from previously transmitted syntax

elements

It is also possible to have both motion predictions from past, or both motion predictions from future.


Transform: types

Each residual macroblock is transformed, quantized and coded

H.264 uses a smaller size transform

H.264 uses three transforms depending on the type of residual data that has to be coded

• a 4x4 transform for the luma DC coefficients in Intra_16x16 macroblocks

• a 2x2 transform for the chroma DC coefficients• a 4x4 transform for all other blocks

Adaptive block size transform mode

Further transforms are (eventually) chosen depending on the motion compensation block size (4x8, 8x4, 8x8, 16x8, etc)


Transform: Order

For a 16x16 Intra mode coded Macroblock

• “-1” Block DC coefficient of each 4x4 luma block• “0-15” Blocks Luma residual blocks• “16-17” Blocks DC coefficients from the Cb and Cr components• “18-25” Bloks Chroma residual blocks

Coding of smooth areas


4x4 residual transform

This transform operates on 4x4 blocks of residual data (labelled 0-15 and 18-25)

Fundamental differences from DCT transform:

• integer transform• fully specified inverse transform mismatch between encoders and

decoders should not occur• requires only additions and shifts• scaling multiplication integrated into the quantizer• transform and quantization can be carried out using 16-bit integer

arithmetic

Reasons

• better prediction method less spatial correlation in the residual• visual benefits: less noise around edges• less computations and a smaller processing wordlength


Transform: Order

The 4x4 DCT of an input array X is given by:

• CXCT “core” 2-D transform• E matrix of scaling factors

d=c/b


Transform: Order

To simplify the implementation of the transform, while keeping it orthogonal, the parameters must be modified

The inverse transform is given by:


Quantization

Requirements

• avoid division and/or floating point arithmetic• incorporate the post- and pre-scaling matrices Ef and Ei described

above

A total of 52 values of Qstep are supported by the standard and these are indexed by a Quantization Parameter, QP.

• Qstep doubles in size for every increment of 6 in QP• Qstep increases by 12.5% for each increment of 1 in QP• QP may be different for luma and chroma (an offset can be signalled

in a Picture Parameter Set)


DC Components Transform

Luma DC coefficient (Intra_4x4 block)

• 4x4 Hadamard transform:

Chroma DC coefficient (any block)

• 2x2 Hadamard transform:


Scanning

The quantized transform coefficients of a block generally are scanned in a zig-zag fashion and transmitted using entropy coding methods

The 2x2 DC coefficients of the chroma component are scanned in raster-scan order and transmitted using entropy coding methods


De-Blocking Filter

Benefits of the application to every decoded macroblock:

• reduced blocking distortion block improved image appearance• better motion-compensated prediction of further frames

Filtering is applied to vertical or horizontal edges of 4x4 blocks:

1) Filter 4 vertical boundaries of the luma component (a,b,c,d)2) Filter 4 horizontal boundaries of the luma component (e,f,g,h)3) Filter 2 vertical boundaries of each chroma component (i,j)4) Filter 2 horizontal boundaries of each chroma component (k,l)


Filtering Choice

The choice of filtering outcome depends on

• boundary strength• gradient across the boundary

BS Rule

4 p or q is intra coded and boundary is a macroblock boundary

3 p or q is intra coded and boundary is not a macroblock boundary

2 neither p or q is intra coded; p or q contain coded coefficients

1 neither p or q is intra coded; neither p or q contain coded coefficients; p and q have different reference frames or a different number of reference frames or different motion vector values

0 neither p or q is intra coded; neither p or q contain coded coefficients;p and q have same reference frame and identical motion vectors no filtering

The filter is “stronger” where there is likely to be significant blocking distortion (high values of BS)


Filter decision

A group of samples is filtered only if BS > 0 and the following conditions issatisfied:

• |p0-q0| < (QP)• |p1-p0| < (QP)• |q1-q0| < (QP) with (QP)< (QP)

Small QPAnything other than a very small gradient across the boundary is likely to be due to image features low (QP) and (QP) value

Large QPblocking distortion is likely to be more significant high (QP) and (QP) value, so that more filtering takes place.


Filter implementation

I. BS=1,2,3

• A 4-tap linear filter is applied with inputs p1, p0, q0 and q1, producing filtered outputs P0 and Q0

• For luma onlyo If |p2-p0| < (QP), a 4-tap linear filter is applied with inputs p2, p1,

p0 and q0, producing filtered output P1

o If |q2-q0| < (QP), a 4-tap linear filter is applied with inputs q2, q1, q0 and p0, producing filtered output Q1

II. BS=4

If |p2-p0|< (QP) and |p0-q0|< (QP) /4• P0 is produced by 5-tap filtering of p2, p1, p0, q0 and q1

• P1 is produced by 4-tap filtering of p2, p1, p0 and q0

• Luma P2 is produced by 5-tap filtering of p3, p2, p1, p0 and q0

else:• P0 is produced by 3-tap filtering of p1, p0 and q1

(the same for qi pixels)


De-blocking: Example

Original Frame

Reconstructed Frame; QP=32; No Filter Reconstructed Frame; QP=32; With Filter

Reconstructed Frame; QP=36; No Filter Reconstructed Frame; QP=36; With Filter


Entropy Coding

Parameters that require to be encoded and transmitted

• Macroblock Type (Prediction method)• QP• Motion data: Reference frame and Motion Vector• Coded block pattern (blocks containing coded coefficients)• Residual Data

H.264/MPEG-4 AVC uses a number of techniques for entropy coding:

• Exp-Golomb codes

All syntax elements except the quantized transform coefficients.

• Context Adaptive Variable Length Coding (CAVLC)

Quantized transform coefficients

• Context Adaptive Binary Arithmetic Coding (CABAC).


Exp-Golomb Codes

Exp-Golomb codes (Exponential Golomb codes) are variable length codes with a regular construction.

• uses a single infinite-extent codeword table• only the mapping to the single codeword table is customized

according to the data statistics

Each codeword is constructed as follows:

[M zeros][1][INFO] codeword length (2M+1) bits

EncodingM = round[log2(code_num+1)]INFO = code_num+1-2M

DecodingCode_num = 2M +INFO-1


Exp-Golomb Codes: Value mapping

Each element v is assigned a type reflecting how data is to be mapped

• Unsigned Exponential ue(v) Maping: code_num = v Used for macroblock type, reference frame index

• Signed Exponential se(v) Mapping

o code_num = 2|v| (v < 0)o code_num = 2|v| - 1(v > 0)

Used for motion vector difference, QP.

• Mapped Exponential me(v) Mapping specified in the standard Used for the Coded block patterns.

• Truncated Exponential te(v)

Each mapping is designed to produce short codewords for frequently occurring values and longer codewords for less common parameter values.

Table for Inter predicted macroblocks: coded_block_pattern indicates which 8x8 blocks in a macroblock contain non-zero coefficients


CAVLC

Method for coding residual, zig-zag ordered transformed blocks

VLC tables for various syntax elements are switched depending on already transmitted syntax elements

CAVLC takes advantage of several characteristics of quantized 4x4 blocks:

• Sparse blocks (containing mostly zeros)

• Highest non-zero coefficients after the zig-zag scan are often sequences of +/-1 (“Trailing 1s” or “T1s”)

• The number of non-zero coefficients in neighbouring blocks is correlated

• The level (magnitude) of non-zero coefficients tends to be higher at the start of the reordered array (near the DC coefficient) and lower towards the higher frequencies


CAVLC encoding

CAVLC encoding of a block of transform coefficients proceeds as follows:

1. Encode the number of coefficients and trailing ones in a 4x4 block

2. Encode the sign of each T1

3. Encode the levels of the remaining non-zero coefficients

4. Encode the total number of zeros before the last coefficient

5. Encode each run of zeros


CAVLC: Steps (1/3)

Step 1: Encode the number of coefficients (TotalCoeffs) and T1s

• TotalCoeffs range from 0 to 16 (in a 4x4 block)• T1s range from 0 to 3 (max value allowed)

Coding: 4 choices for look-up table (coeff_token)

• Num_VCL0 biased towards small numbers of coefficients• Num_VCL1 biased towards medium numbers of coefficients• Num_VCL2 biased towards higher numbers of coefficients• Num_FLC fixed 6-bit length code

The choice depends on the number of non-zero coefficients in upper and left-hand previously coded blocks NU and NL (context adaptivity)

• U,L available N=(NU+NL)/2• U available N=NU

• L available N=NL

• U,L unavailable N=0


CAVLC: Steps (2/3)

Step 2: Encode the sign of each T1.

For each T1 up to three, a single bit encodes the sign (0=+, 1=-). Encoding is in reverse order, starting with the highest-frequency T1

Step 3: Encode the levels of the remaining non-zero coefficients.

The level (sign and magnitude) of each remaining non-zero coefficient in the block is encoded in reverse order

Encoding VLC table depends on successive coded level (context adaptivity)

1. Initialise the table to Level_VLC0

2. Encode highest-frequency coefficient

3. If magnitude is larger than a threshold, move up to the next VLC table


CAVLC: Steps (2/3)

Step 4: Encode the total number of zeros before the last coefficient

Coding with a VLC of the number of all zeros preceding the highest non-zero coefficient.

Step 5:Encode each run of zeros

The number of zeros preceding each non-zero coefficient (run_before) is encoded in reverse order.

The VLC for each run of zeros is chosen depending on

• the number of zeros that have not yet been encoded (ZerosLeft)• value of run_before parameter.


CAVLC: example (1/2)

4x4 Block

Reordered Block: 0,3,0,1,-1,-1,0,1,0,0,0,0,0,0,0,0 TotalCoeff = 5; TotalZeros=3; T1s =3 (max value)

Encoding

1

2

3

4

5

Transmitted bitstream 000010001110010111101101 (24 bits)


CAVLC: example (2/2)

Decoding

Input string: 000010001110010111101101Values added to the output array at each stage are underlined

The decoder has inserted two zeros; however, TotalZeros is equal to 3 and so another 1 zero is inserted before the lowest coefficient, making the final output array: 0,3,0,1,-1,-1,0,1,…


CABAC

CABAC uses arithmetic coding

effective use of probability models of occurrence of symbols

Particularly beneficial for symbol probabilities greater than 0.5

The use of adaptive codes permits adaptation to non-stationary symbol statistics

easily adapt to changing statistical characteristics of the data to be coded

Context modelling: The statistics of already-coded syntax elements are used to estimate the conditional probabilities

CABAC provides a reduction in bit-rate between 5% to 15% over CAVLC, when coding TV signals at the same quality.


CABAC basis

Encoding with CABAC consists of three stages

• Binarization needed for syntax elements that are non-binary valued

• Context modeling a model is selected such that the choice may depend on previous

encoded syntax elements or bins

• Adaptive binary arithmetic coding coding of bin sequence with updating of probabilities takes place


Binarization

Requirements for a successful context coding method

• fast and accurate estimation of conditional probabilities• the computational complexity must be kept at a minimum

“Pre-processing” step to reduce the alphabet size of the syntax elements

The design of binarization schemes in CABAC (mostly) relies on a few basic code trees

• Unary or Truncated Unary code x>0 x “1” bits plus a terminating “0” bit (unary) 0<x<S for x=S the terminating “0” bit is neglected (truncated)• Exp-Golomb code codes are constructed by a concatenation of a prefix and a suffix

code word• Fixed-length code a finite alphabet of values of the corresponding syntax element is

assumed.


Context Modeling

A model probability distribution is assigned to the given symbols, to drive the actual coding engine to generate a sequence of bits as a coded representation of the symbols

• Define a modeling function F:T C operating on a set T of past symbols

• For each symbol x to be coded, a conditional probability p(x|F(z)) is estimated according to the already coded neighboring symbols z T

• After encoding x, the probability model is updated with the value of the encoded symbol x

Context template consisting of two neighboring syntax elements A and B to the left and on top of the current syntax element C


Profiles and LevelsProfile Defines a set of coding tools or algorithms that can be used in generating a conforming bit-stream

LevelPlaces constraints on certain key parameters of the bitstream

Decoders conforming to a profile must support all features in that profile

In H.264/AVC, three profiles are defined (plus extensions)• Baseline• Main• Extended Profile• Fidelity Range Extensions: four HighProfile versions (High, High 10,

High 4:2:2, and High 4:4:4), for high quality uses

Fifteen levels are defined specifying upper limits for• picture size• decoder-processing rate • size of the multi-picture buffers

• video bit rate• video buffer size


Profile features

Baseline profile supports all features in H.264/AVC except the following two feature sets:

• Set 1o B sliceso Weighted predictiono CABACo Field codingo Picture or macroblock adaptive switching between frame and field coding

• Set 2o SP/SI slices

Main profile supports Set 1, and does not support the FMO and redundant pictures features

Extended Profile supports all features except for CABAC.


Profile Map


Application Areas

Conversational services• operate typically below 1 Mbps; low latency requirements• Baseline profile for following application

• H.320 conversational services with circuit-switched ISDN-based video conferencing• 3GPP conversational H.324/M services• H.323 conversational services over the Internet with best effort IP/RTP protocols.• 3GPP conversational services using IP/RTP for transport and SIP for session set-up.

Entertainment video applications• operate between 1-8 Mbps; moderate latency (0.5 to 2 seconds)• Main profile for following application

• Broadcast via satellite, cable, terrestrial, or DSL• DVD for standard and high-definition video• Video on demand via various channels.

Streaming services• operate at 50 kbps–1.5 Mbps; latency of 2 or more seconds• Baseline or Extended profile; differences based on the use for wired

or wireless environments as follows:• 3GPP streaming using IP/RTP for transport and RTSP Baseline profile• Streaming over the wired Internet using IP/RTP protocol and RTSP for session set-up

Extended profile


Differences


Complexity of Codec Design

Codec design includes relaxation of traditional bounds on complexity (memory & computation) – rough guess 2-3x decoding power increase relative to MPEG-2, 3-4x encoding

Problem areas:• Smaller block sizes for motion compensation (cache access issues)• Longer filters for motion compensation (more memory access)• Multi-frame motion compensation (more memory for reference frame

storage)• More segmentations of macroblock to choose from (more searching

in the encoder)• More methods of predicting intra data (more searching)• Arithmetic coding (adaptivity, computation on output bits)


Comparison (1/2)

27

28

29

30

31

32

33

34

35

36

37

38

39

0 50 100 150 200 250

Bit-rate [kbit/s]

Foreman QCIF 10Hz

MPEG-2

H.263

MPEG-4

JVT/H.264/AVC

Qu

ality

Y-P

SN

R [

dB

]


Comparison (2/2)

Tempete CIF 30Hz

25

26

27

28

29

30

31

32

33

34

35

36

37

38

0 500 1000 1500 2000 2500 3000 3500

Bit-rate [kbit/s]

Qu

ality

Y-P

SN

R [

dB

]

MPEG-2

H.263

MPEG-4

JVT/H.264/AVC


Test Set Results for Perceptual Quality

Informal perceptual tests

At the same PSNR, people generally prefer JVT

• Small motion compensation block size (breaks up block structure)• Small transform block size (breaks up block structure, reduces

ringing)• In-loop deblocking filter

By how much?

• Needs further study• No rigorous testing reported• 10-15% might be a good guess


References

IEEE Transactions on Circuits and Systems for Video Technology (2003):Special Issue on the H.264/AVC Video Coding Standard

Signal Processing: Image Communication (2004):Video coding using the H.264/MPEG-4 AVC compression standard

http://www.vcodex.com/h264.html

università degli studi roma tre – introduzione alla codifica h.264/avc università degli studi...

Documents

avc introduction

avc history

avc applications

avc technology

avc terminology

avc evolution

standard slide

avc versions