MPEG-FAQ: multimedia compression [2/6]

MPEG-FAQ: multimedia compression [2/6]

Post by Frank Gadega » Thu, 02 Mar 1995 03:41:34

Archive-name: mpeg-faq/part2
Last-modified: 1994/08/22
Version: v 3.2 94/08/22
Posting-Frequency: bimonthly

BEGIN -------------------- CUT HERE --------------------- 2/6
   for example, the MPEG-1 style sequence_header() is followed by
   sequence_extension() which is exclusive to MPEG-2. Some extension
   headers are specific to MPEG-2 profiles. For example,
   sequence_scalable_extension() is not allowed in Main Profile.

   A simple program need only scan the coded bistream for byte-aligned
   start codes to determine whether the stream is MPEG-1 or MPEG-2.

Q. What is the precision of MPEG samples?
A. By definition, MPEG samples have no more and no less than 8-bits
   uniform sample precision (256 quantization levels).  For luminance
   (which is unsigned) data, black corresponds to level 0, white
   is level 255. However, in CCIR recommendation 601 chromaticy, levels
   0 through 14 and 236 through 255 are reserved for blanking signal
   excursions. MPEG currently has no such clipped excursion restrictions.

Q. Is it MPEG-2 (arabic numbers) or MPEG-II (roman)?

A. Committee insiders most often use the arabic notation with the
   hyphen, e.g. MPEG-2.  Only the most retentive use the official
   designation: Phase 2.   In fact, M.P.E.G. itself is a nickname.  The
   official name is: ISO/IEC JTC1 SC29 WG11.  The militaristic lingo has
   so far managed to keep the enemy (DVI) confused and out of the picture.

   ISO:  International Organization for Standardization
   IEC:  Interntional Electrotechnical Commission
   JTC1: Joint Technical Committee 1
   SC29: Sub-committee 29
   WG11: Work Group 11  (moving pictures with... uh, audio)

Q. Why MPEG-2?  Wasn't MPEG-1 enough?

A. MPEG-1 was optimized for CD-ROM or applications at about 1.5 Mbit/sec.
   Video was strictly non-interlaced (i.e. progressive).  The international
   co-operation had executed so well for MPEG-1, that the committee began to
   address applications at broadcast TV sample rates using the CCIR 601
   recommendation (720 samples/line by 480 lines per frame by 30 frames per
   second... or about 15.2 million samples/sec including chroma) as the

   Unfortunately, today's TV scanning pattern is interlaced.  This
   introduces a duality in block coding:  do local redundancy areas
   (blocks) exist exclusively in a field or a frame...
   (or a particle or wave) ?  The answer of course is that some blocks
   are one or the other at different times, depending on motion activity.  

   The additional man years of experimentation and implementation between
   MPEG-1 and MPEG-2 improved the method of block-based transform coding.

Q. How do MPEG and JPEG differ?

A. The most fundamental difference is MPEG's use of block-based motion
   compensated prediction (MCP)---a general method falling into the
   temporal DPCM category.  

   The second most fundamental difference is in the target application.
   JPEG adopts a general purpose philosophy: independence from color space
   (up to 255 components per frame) and quantization tables for each
   component.  Extended modes in JPEG include two sample precisions (8 and
   12 bit sample accuracy), combinations of frequency progessive, spatially
   progressive, and amplitude progressive scanning modes. Color independence
   is made possible thanks to downloadable Huffman tables.

   Since MPEG is targeted for a set of specific applications, there is
   only one color space (4:2:0 YCbCr), one sample precision (8 bits), and
   one scanning mode (sequential). Luminance and chrominance share
   quantization tables. The range of sampling dimensions are more limited
   as well.  MPEG adds adaptive quantization at the macroblock (16 x 16 pixel
   area) layer.  This permits both smoother bit rate control
   and more perceptually uniform quantization throughout the picture and
   image sequence.   Adaptive quantization is part of the JPEG-2 charter.
   MPEG variable length coding tables are non-downloadable, and are
   therefore optimized for a limited range of compression ratios
   appropriate for the target applications.

   The local spatial decorrelation methods in MPEG and JPEG are very similar.
   Picture data is block transform coded with the two-dimensional orthanormal
   8x8 DCT. The resulting 63 AC transform coefficients are mapped in a
   zig-zag pattern to statistically increase the runs of zeros. Coefficients
   of the vector are then uniformily scalar quantized, run-length coded, and
   finally the run-length symbols are variable length coded using a
   cannonical (JPEG) or modified Huffman (MPEG) scheme.  Global frame
   redundancy is reduced by 1-D DPCM of the block DC coefficients, followed
   by quantization and variable length entropy coding.

            MCP                   DCT                    ZZ               Q
       Frame -> 8x8 spatial block -> 8x8 frequency block -> Zig-zag scan ->

                    RLC                  VLC
       quanitzation -> run-length coding -> variable length coding.

   The similarities have made it possible for the development of hard-wired
   silicon that can code both standards.  Even microcoded architectures can
   better optimize through hardwired instruction primitives or functional
   blocks. There are many additional minor differences. They include:

     1. DCT and quantization precision in MPEG is 9-bits since the macroblock
        difference operation expands the 8-bit signal precision by one bit.

     2. Quantization in MPEG-1 forces quantized coefficients to become
        odd values (oddification).

     3. JPEG run-length coding produces run-size tokens (run of zeros,
        non-zero coefficient magnitude) whereas MPEG produces fully
        concatenated run-level tokens that do not require magnitude
        differential bits.

     4. DC values in MPEG-1 are limited to 8-bit precision (a constant
        stepsize of 8), whereas JPEG DC precision can occupy all possible
        11-bits.  MPEG-2, however, re-introduced extra DC precison.

Q. What happened to MPEG-3?

A. MPEG-3 was to have targeted HDTV applications with sampling dimensions
   up to 1920 x 1080 x 30 Hz and coded bitrates between 20 and 40 Mbit/sec.
   It was later discovered that with some (compatible) fine tuning, MPEG-2
   and MPEG-1 syntax worked very well for HDTV rate video.  The key is
   to maintain an optimal balance between sample rate and coded bit rate.

   Also, the standardization window for HDTV was rapidly closing.  Europe
   and the United States were on the brink of committing to analog-digital
   subnyquist hybrid algorithms (D-MAC, MUSE, et al).   European all-digital
   projects such as HD-DIVINE and VADIS demonstrated better picture quality
   with respect to bandwidth using the MPEG syntax.  In the United States, the
   Sarnoff/NBC/Philips/Thomson HDTV consortium had used MPEG-1 syntax from
   the beginning, and with the exception of motion artificats (due to
   limited search range in the encoder), was deemed to have the best picture
   quality of all three digital proponents.

   HDTV is now part of the MPEG-2 High-1440 Level and High Level toolkit.

Q. What is MPEG-4?
A. MPEG-4 targets the Very Low Bitrate applications defined loosly
   as having sampling dimensions up to 176 x 144 x 10 Hz and coded
   bit rates between 4800 and 64,000 bits/sec.   This new standard would
   be used, for example, in low bit rate videophones over analog
   telephone lines.  

   This effort is in the very early stages.  Morphology, fractals, model
   based, and anal retentive block transform coding are all in the offering.
   MPEG-4 is now in the application identification phase.

Q. Where can I get a copy of the latest MPEG-2 draft?
A. Contact your national standards body (e.g. ANSI Sales in NYC for the U.S.)

Q. What is the latest working drafts of MPEG-2 ?
A. The latest versions of video (version 4), and systems were produced at
   the Brusells meeting (September 10, 1993).  The latest audio working
   draft was produced in New York (July 1993).

   MPEG-2 Video, Audio, and Systems will reach CD at the November 1994
   Seoul, Korea meeting.

Q. What is the latest version of the MPEG-1 documents?
A. Systems (ISO/IEC IS 11172-1), Video (ISO/IEC IS 11172-2), and Audio
   (ISO/IEC IS 11172-3) have reached the final document stage.  Part 4,
   Conformance Testing, is currently a CD.

Q. What is the evolution of standard documents?
A. In chronological order:

   New Proposal (NP)
   Working Draft (WD)
   Committee Draft (CD)
   Draft International Standard (DIS)
   International Standard (IS)

Q. When will an MPEG-2 decoder chip be available?
A. Several chips will be sampling in late 1993.  For reasons of economy
   and scale in the cable TV application, all are single-chip (not including
   DRAM and host CPU/controller) implementations.
   They are:

  SGS-Thomson STi-3500
        first MPEG-2 chip on market
        multi-tap binary horizontal sample rate convertor.
        pan & scanning support for 16:9
        requires external, dedicated microcontroller (8 bit)
        8-bit data bus, no serial data bus.

  LSI Logic L64112 successor (pin compatible)
        serial bus, 15 Mbit coded throughput.
        smaller pin-count version due soon.

  C-Cube CL-950 successor (?)

  In 1994, we can look forward to:

  Pioneer single-chip MPEG-2 successor to CD-1100 MPEG-1 chip set.
  IBM single-chip decoder.

Q. Are there single chip MPEG encoders?

A. Yes, the C-Cube CL-4000 is the only single-chip, real-time encoder
   that can process true MPEG-1 SIF rate video.

   Single chip for +/- 15 pel motion estimation at SIF rates (352x240x30 Hz)
   Two chips for +/- 32 pel at SIF rates (hierarchical)
   5 or 6 chips for MPEG-2 at CCIR 601 rates (704 x 480 x 30 Hz)
   Highly microcoded architecture.
   Can code both H.261 and JPEG.
   Implements high picture quality microcode programs.
   [more details from CICC'93 and HotChips '93 conference to be included]

   IBM and SGS-Thomson plan to introduce more hard-wired, multichip
   solutions in 1994.

Q. What about MPEG-1 decoder chips?

A. By implication of MPEG-2 Conformace requirements, all MPEG-2 decoders are
   required to decode MPEG-1 bitstreams as well. These chips, however, are
   strictly MPEG-1:

        C-Cube CL-450           SIF rates. Single-chip.  Has on-board CPU.

        SGS-Thomson 3400        SIF rates. Single-chip.  Hardwired.

        Motorola MCD250         SIF rates. Single-chip.  

        LSI 641172              CCIR 601 rates. Single-chip.  Systems
                                packet decoder on-chip.

Q. What about audio chips?
A. To date, only Layer I and Layer II have been implemented in dedicated
   (ASIC) silicon:

  Motorola MCD260

  Texas Instruments TI 320AV110
        hardwired with systems parsing)                      
        operates in free format (arbitrary sample rate)
        120 pin PQFP package
        Serial data port
        Part of technology exchange with C-Cube

  LSI Logic L64111
        hardwired w/CPU with on-chip systems parsing.
        Serial data port                        
        100-pin PQFP              


  Crystal Semiconductor CS4920
        on-chip, 2 channel 16-bit digital-to-analog convertor (DAC)
        16 MIPS, 24-bit DSP
        programmable clock manager
        44-pin PLCC package
        Programmable architecture.  For example, can download Layer II
          MPEG-1 audio or Dolby AC-2
        $38 each in large quantities

Dolby AC-3
        MPEG NY disclosure
        claimed to be less computationally intensive
        Zoran, GI working on own DSP-like dedicated chips.

Q. Will there be an MPEG video tape format?

A. There is a consortium of companies (Philips, JVC, Sony, Matushista,
   et al) developing a metal particle based 6 milimeter consumer digital
   video tape format. It will initially use more JPEG-like independent
   frame compression for cheap encoding of source analog (NTSC, PAL)
   video.  The consequence of course is less efficient use of bandwidth (
   25 Mbit/sec for the same quality acheived at 6 Mbit/sec with MPEG).
   Pre-compressed video from broadcast sources will be directly recorded
   to tape and "passed-through" as a coded bitstream to the video
   decompression "box" upon playback.

Q. What do B-frames buy you?
A. Since bi-directional marcoblock predictions are an average of two maroblocks blocks,
   noise is reduced at low bit rates.  At nominal MPEG-1 video (352 x 240 x 30, 1.15
   Mbit/sec) rates, it is said that B-frames improves SNR by as much as 2 dB.
   (0.5 dB gain is usually considered worth-while in MPEG). However, at higher
   bit rates, B-frames become less useful since they inherently do not contribute
   to the progressive refinement of an image sequence (i.e.not used as
   prediction by subsequent coded frames).  Regardless, B-frames are still
   politically controversial.

Q. Why do some people hate B-frames?
A. Computational complexity, bandwidth, delay, and picture buffer size are
   the four B-frame Pet Peeves. Computational complexity is increased since
   a some macroblock modes require averaging between two macroblocks.  
   Worst case, memory bandwidth is increased an extra 16 MByte/s (601
   rate) for this extra prediction. An extra picture buffer is needed to
   store the future prediction reference (bi-directionality).  Finally,
   extra delay is introduced in encoding since the frame used for backwards
   prediction needs to be transmitted to the decoder before the intermediate
   B-pictures can be decoded and displayed.

   Cable television (e.g. General Instruments) have been particularly
   adverse to B-frames since the extra picture buffer pushes the decoder
   DRAM memory requirements past the magic 8-Mbit (1 Mbyte) threshold into the
   realm of 16 Mbits (2 MByte) for CCIR 601 frames (704 x 480), yet not for
   lowly 352 x 480. However, cable does not realize that DRAM does not come
   in convenient high-volume (low cost) 8-Mbit packages as 16-Mbit does.  In
   a few years, the cost differences between 16 Mbit and 8 Mbit will become
   insignificant compared to the gain in compression.  For the time being,
   cable boxes will start with 8-Mbit and allow future drop-in upgrades to
   16-Mbit.  The early market success of B-frames seem to have been
   determined by a fire at a Japanese chemical plant.

Q. How do MPEG and H.261 differ?
A. H.261 was targeted for teleconferencing applications where motion
   is naturally more limited. Motion vectors are restricted to a range of
   +/- 15 pixels.  Accuracy is reduced since H.261 motion vectors are
   restricted to integer-pel accuracy.  Other syntactic differences
   include: no B-pictures, different quantization method.

   H.261 is also known as P*64. "P" is an integer number meant to
   represent multiples of 64kbit/sec.  In the end, this nomenclature
   probably won't be used as many services other than video will adopt the
   philosophy of arbitrary B channel (64kbit) bitrate scalability.

Q. Is H.261 the de facto teleconferencing standard?

A. Not exactly.  To date, about seventy percent of the industrial  
   teleconferencing hardware market is controlled by PictureTel of Mass.  
   The second largest market controller is Compression Labs of Silicon
   Valley.  PictureTel hardware includes compatibility with H.261 as a
   lowest common denominator, but when in comminication with other
   PictureTel hardware, it can switch to a mode superior at low bit rates
   (less than 300kbits/sec). In fact, over 2/3 of all teleconfercing is done
   at two-times switched 56 channel (~P = 2) bandwidth.  Long distance ISDN
   ain't cheap.  In each direction, video and audio are coded at an
   aggregate of 112 kbits/sec (2*56 kbits/sec).

   The PictureTel proprietary compression algorithm is acknowledged to
   be a combination of spatial pyramid, lattice vector quanitzer, and an
   unidentified entropy coding method.  Motion compensation is considerably
   more refined and sophisticated than the 16x16 integer-pel block method
   specified in H.261.

   The Compression Labs proprietary algorithm also offers significant
   improvement over H.261 when linked to other CLI hardware.

   Currently, ITU-TS (International Telecommunications Union--Teleconferencing
   Sector), formerly CCITT, is quietly defining an improvement to H.261 with
   the participation of industry vendors.

Q. Where will be see MPEG in everyday life?
A. Just about wherever you see video today.

   DBS (Direct Broadcast Satellite)
     The Hughes/USSB DBS service will use MPEG-2 video and audio.  Thomson
     has exclusive rights to manufacture the decoding boxes for the first
     18 months of operation.  No doubt Thomson's STi-3500 MPEG-2 video
     decoder chip will be featured.

     Hughes/USSB DBS will begin service in North America in April 1994.
     Two satellites at 101 degrees West will share the power requirements  
     of 120 Watts per 27 MHz transponder. Multi-source channel rate
     control methods will be employed to optimally allocate bits between
     several programs on one data carrier. An average of 150 channels are  

   CATV (Cable Television)
     Despite conflicting options, the the cable industry has more or less
     settled on MPEG-2 video.  Audio is less than settled. For example,
     General Instruments (the largest U.S. consumer cable set-top box
     manufacturer) have announced the planned use of the Dolby AC-3
     audio algorithm.

     The General Instruments DigiCipher I video syntax is similar to MPEG-2
     syntax but uses smaller macroblock predictions and no B-frames.  The
     DigiCipher II specification will include modes to support both the GI
     and full MPEG-2 Video Main Profile syntax.  Services such as HBO will
     upgrade to DigiCipher II in 1994.

     The U.S. Grand Alliance, a consortium of companies that formely competed
     for the U.S. terrestrial HDTV standard,  have already agreed to use
     the MPEG-2 Video and Systems syntax---including B-pictures. Both interlaced
     (1440 x 960 x 30 Hz) and progressive (1280 x 720 x 60 Hz) modes will
     be supported. The Alliance must then settle upon a modulation (QAM,
     VSB, OFDM), convolution (MS or Viterbi), and error correction (RSPC, RSFC)

     In September 1993, the consortium of 85 European companies signed an
     agreement to fund a project known Digital Video Broacasting (DVB) which
     will develop a standard for cable and terrestrial transmission by the
     end of 1994. The scheme will use MPEG-2.  This consortium has put the
     final nail in the coffin of the D-MAC scheme for gradual migration
     towards an all-digital, HDTV consumer transmission standard. The only
     remaining analog or digital-analog hybrid system left in the world is
     NHK's MUSE (which will probably be axed in a few years).

Q. What did MPEG-2 add to MPEG-1 in terms of syntax/algorithms ?
A. Here is a brief summary:

  Sequence layer:
  More aspect ratios.  A minor, yet neccessary part of the syntax.

  Horizontal and vertical dimensions are now required to be a multiple of
  16 in frame coded pictures, and the vertical dimension must be a multiple
  of 32 in field coded pictures.

  4:2:2 and 4:4:4 macroblocks were added in the Next profiles.

  Syntax can now signal frame sizes as large as 16383 x 16383.

  Syntax signals source video type (NTSC, PAL, SECAM, MAC, component) to
  help post-processing and display.

  Source video color primaries (609, 170M, 240M, D65, etc.) and opto-
  electronic transfer characteristics (709, 624-4M, 170M etc.) can be

  Four scalable modes [see scalable section below]

  Picture layer:
  All MPEG-2 motion vectors are half-pel accuracy.

  DC precision can be user-selected as 8, 9, 10, or 11 bits.

  Concealment motion vectors were added to I-pictures in order to
  increase robustness from bit errors since I pictures are the most
  critical and sensitive in a group of pictures.

  A non-linear macroblock quantization factor that results in a more
  dynamic step size range, from 0.5 to  56, than in MPEG-1 (1 to 32).

  New Intra-VLC table for dct_next_coefficient (AC run-level events)
  that is more geared towards I-frame probability distribution.  EOB
  is 4 bits.  The old tables are still included.

  Alternate scanning pattern that (supposedly) improves entropy coding
  performance over the original Zig-Zag scan used in H.261, JPEG, and
  MPEG-1.  The extra scanning pattern is geared towards interlaced

  Syntax to signal 3:2 pulldown process (repeat_field_first flag)

  Syntax flag to signal chrominance post processing type (4:2:0 to
  4:2:2 upsampling conversion)

  Progressive and interlaced frame coding

  Syntax to signal source composite video characteristics useful in
  post-processing operations. (v-axis, field sequence, sub_carrier,
  phase, burst_amplitude, etc.)

  Pan & scanning syntax that tells decoder how to, for example, window a
  4:3 image within a wider 16:9 aspect ratio image.  Vertical pan offset
  has 1/16th pixel accuracy.

  Macroblock layer:
  Macroblock stuffing is now illegal in MPEG-2 (hurray!!)

  Two line modes (interlaced and progressive) for DCT operation.  

  Now only one run-level escape code code (24-bits) instead of
  the single (20-bits) and double escape (28-bits) in MPEG-1.

  Improved mismatch control in quantization over the original oddification
  method in MPEG-1.  Now specifies adding or subtracting one to the
  63rd AC coefficient depending on parity of summed quantized coefficients.

  Many additional prediction modes (16x8 MC, field MC, Dual Prime)
  and, correspondingly, macroblock modes.

  Overall, MPEG-2's greatest compression improvements over MPEG-1 are:  
  prediction modes, Intra VLC table, DC precision, non-linear macroblock
  quant.  Implementation improvements, well,.. uh... macroblock stuffing
  was eliminated.

Q. What are the scalable modes of MPEG-2?
A. Scalable video is permitted only in the Main+ and Next profiles.
   Currently, there are four scalable modes in the MPEG-2 toolkit.
   These modes break MPEG-2 video into different layers (base, middle,
   and high layers) mostly for purposes of prioritizing video data.  For
   example, the high priority channel (bitstream) can be coded with a
   combination of extra error correction information and decreased bit
   error (i.e. higher Carrier-to-Noise ratio or signal strength) than
   the lower priority channel.  

   Another purpose of scalablity is complexity division.  For example,
   in HDTV, the high priority bitstream (720 x 480) can be decoded
   under noise conditions were the lower priority (1440 x 960) cannot.
   This is "graceful" degradation. By the same division however, a
   standard TV set need only decode the 720 x 480 channel, thus requiring
   a less expensive decoder than a TV set wishing to display 1440 x 960.
   This is simulcasting.

   A brief summary of the MPEG-2 video scalability modes:
   [better descriptions in installment 3]  

   Spatial Scalablity-- Useful in simulcasting, and for feasible software
    decoding of the lower resoultion, base layer.  This spatial domain
    method codes a base layer at lower sampling dimensions (i.e. "resolution")
    than the upper layers.  The upsampled reconstructed lower (base) layers
    are then used as prediction for the higher layers.  

   Data Partitioning-- Similar to JPEG's frequency progressive mode, only
    the slice layer indicates the maximum number of block transform
    coefficients contained in the particular bitstream (known as the
    "priority break point").  Data partitioning is a frequency domain method
    that breaks the block of 64 quantized transform coefficients into two
    bitstreams.  The first, higher priority bitstream contains the more
    critical lower frequency coefficients and side informations (such as DC
    values, motion vectors). The second, lower priority bitstream carries
    higher frequency AC data.

   SNR Scalability-- Similar to the point transform in JPEG, SNR scalability
    is a spatial domain method where channels are coded at identical sample
    rates, but with differing picture quality (through quantization step sizes).  
    The higher priority bitstream contains base layer data that can be added
    to a lower priority refinement layer to construct a higher quality picture.

   Temporal Scalability--- A temporal domain method useful in, e.g.,
    stereoscopic video.  The first, higher priority bitstreams codes video
    at a lower frame rate, and the intermediate frames can be coded in a
    second bitstream using the first bitstream reconstruction as prediction.  
    In sterescopic vision, for example, the left video channel can be
    prediction from the right channel.

   Other scalability modes were experimented with in MPEG-2 video (such as
   Frequency Scalability), but were eventually dropped in favor of methods
   that demonstrated similar quality and greater simplicity.

Q. What is all the fuss with cositing of chroma components?
A. It is important to properly co-site chroma samples, otherwise chroma
   shifting may result.  
   [insert more details in installment 3]

Q. What is the reasoning behind MPEG syntax symbols?
A. Here are some of the Whys and Wherefores of MPEG symbols:

  Start codes
  These 32-bit byte-aligned codes provide a mechanism for cheaply searching
  coded bitstreams for commencment of various layers of video without having
  to actually parse or decode.  Start codes also provide a mechanism for
  resynchronization in the presense of bit errors.

  Coded block pattern (CBP --not to be confused with Constrained Parameters!)
  When the frame prediction is particularly good, the displaced
  frame differencene (DFD, or prediction error) tends to be small, often
  with entire block energy being reduced to zero after quantization.  This
  usually happens only at low bit rates.  Coded block patterns prevent
  the need for transmitting EOB symbols in those zero coded blocks.

  Each intra coded block has a DC coefficient.  Inter coded blocks
  (prediction error or DFD) naturally do not since the prediction error
  is the first derivative of the video signal. With coded block patterns
  signalling all possible non-coded block patterns, the dct_coef_first
  mechanism assigns a different meaning to the VLC codeword that would
  otherwise represent EOB as the first coefficient.

  End of Block
  Saves unecessary run-length codes.  At optimal bitrates, there tends to be
  few AC coefficients concentrated in the early stages of the zig-zag vector.
  In MPEG-1, the 2-bit length of EOB implies that there is an average of only
  3 or 4 non-zero AC coefficients per block.  In MPEG-2 Intra (I) pictures,
  with a 4-bit EOB code, this number is between 9 and 16 coefficients.
  Since EOB is required for all coded blocks, its absense can signal that a
  syntax error has occurred in the bitstream.

  Macroblock stuffing
  A genuine pain for VLSI implementations, macroblock stuffing was introduced
  to maintain smoother, constant bitrate control in MPEG-1. However, with
  normalized complexity measures and buffer management performed on a
  a priori (pre-frame, pre-slice, and pre-macroblock) basis in the MPEG-2
  encoder test model, the need for such localized smoothing evaportated.  
  Stuffing can be acheived through virtually unlimited slice start code
  padding if required. A good rule of thumb: if you find yourself often
  using stuffing more than once per slice, you probably don't have a very
  good rate control algorithm.  Anyway, marcoblock stuffing is now illegal in

  MPEG's modified Huffman VLC tables
  The VLC tables in MPEG are not Huffman tables in the true sense of
  Huffman coding, but are more like the tables used in Group 3 fax.
  They are entropy constrained, that is, non-downloadable and optimized
  for a limited range of bit rates (sweet spots).  With the acception of
  a few codewords, the larger tables were carried over from the H.261
  standard of 1990.  MPEG-2 added an "Intra table".  Note that the
  dct_coefficient tables assume positive/negative coefficient pmf symmetry.

Q. What is the TM rate control and adaptive quantization technique ?
A. Test model was not by any strech of the imagination meant to
   be the show-stopping, best set of algorithm.  It was designed to
   excersize the syntax, verify proposals, and test the *relative*
   performance of proposals in a way that could be duplicated
   by co-experimentors in a timely fashion.  Otherwise there would
   be more endless debates about model interpretation than actual
   time spent in verification.

  [MPEG-2 Test model is frozen as v5b]

  The MPEG-2 Test Model (TM) rate control method offers a dramatic
  improvement to the Simulation Model (SM) method used for MPEG-1.  TM's
  improvements are due to more sophistication pre-analysis and post-analysis

  Rate control and adaptive quantization are divided into three steps:

  Step One:       Bit Allocation

    In Complexity Estimation, the global complexity measures assign relative
    weights to each picture type.  These weights (Xi, Xp, Xb) are reflected
    by the typical coded frame size of I, P, and B pictures (see typical frame
    size section). I pictures are assigned the largest weight since they have
    the greatst stability factor in an image sequence.  B pictures are assigned
    the smallest weight since B data does not propogate into other frames
    through the prediction process.

    Picture Target Setting allocates target bits for a frame based on
    the frame type and the remaining number of frames of that same
    type in the Group of Pictures (GOP).

Step Two:       Rate Control

        Rate control attempts to adjust bit allocation if there is
        significant difference between the target bits (anticipated
        bits) and actual coded bits for a block of data.

        [more detail in installment 3]

Step Three:     Adaptive Quantization

        Recomputes macroblock quantization factor according to
        activity of block against the normalized activity of the

        The effect of this step is to roughly assign a constant number
        of bits per macroblock (this results in more perceptually uniform
        picture quality).

        [more detail in installment 3]

Q. How would you explain MPEG to the data compression expert?
A. MPEG video is a block-based video scheme
   Local decorrelations via DCT-Q-VLC hybrid
   Dead-zone quanitizer
   DFD: quantized prediction error
   [etc.  More in installment 3]

Q. What are the implementation requirements?
A. MPEG pushes the limit of economical VLSI technology (but you get
   what you pay for in terms of picture quality or compaction efficiency)

   Video                Typical decoder     Total    DRAM bus width
   Profile              transistor count    DRAM     @ speed
   ------------         ----------------    -------  -------------------
   MPEG-1 CPB           0.4 to .75 million   4 Mbit  16 bits @ 80 ns
   MPEG-1 601           0.8 to 1.1 million  16 Mbit  64 bits @ 80 ns
   MPEG-2 MP@ML         0.9 to 1.5 million  16 Mbit  64 bits @ 80 ns
   MPEG-2 MP@High1440     2 to   3 million  64 Mbit  N/A

   70 or 80ns DRAM speed is a measure of the shortest period in which
   words can be transfered across the bus.  In the case of MPEG-1 SIF,
   80ns implies (1/80ns)(16bits) or about 25 MBytes/sec of bandwidth.
   Lack of cheap memory (DRAM) utilization is where the original DVI
   algorithm made a costly mistake.  DVI required expensive VRAM/SRAM
   chips (a static RAM transistor requires 6 transistors compared to
   1 transistor for DRAM).    Fast page mode DRAM (which has slower
   throughput than SRAM and requires near-contiguous address mapping)
   is viable for MPEG due almost exclusively to the block nature of
   the algorithm and syntax (DRAM memory locations are broken into
   rows and columns).

Q. Is exhuastive search "optimal" ?
A. Definately not in the context of block-based MCP.   Since one motion
   vector represents the prediction of 256 pixels, divergent pixels within
   the macroblock are misrepresented by the "global" vector.  This leads
   back to the general philosophy of block-based coding as an approximation
   technique.  Exhuastive search may find blocks with the least distortion
   (displaced frame difference) but will not produce motion vectors with
   the least entropy. [more details later]

Q. What is a good motion estimation method, then?
   When shopping for motion vectors, the three basic characteristics are:
   Search range, search pattern, and matching criteria.  Search pattern
   has the greatest impact on finding the best vector.   Hierarchical
   search patterns first find the best match between downsampled images of
   the reference and target pictures and then refine the vector through
   progressively higher resolutions.  Hierarchical patterns are less
   likely to be confused by extremely local distortion minimums as being
   a best match.

   [Accuracy vs. Ambiguity]

   [Some ways of solving problem (Gary Sullivan--ICASSP '93), but not
   syntacitally compatible].

   [motion vector pre-frame search, motion vector refinement, etc.
    in installment 3]

Q. What is MPEG 1.5 and MPEG++ ?
A. MPEG-1.5 was not exactly a  proprietary twist in terms of syntax,
   but operating parameters.  Again, people (erronously) consider MPEG-1
   to be limited to SIF rates (352 x 240 x 30 Hz). After interrogation,
   most MPEG 1.5 proponents will confess that MPEG 1.5 is simply MPEG-1 at
   CCIR 601 rates (704 x 480 x 30 Hz) and that it may or may not include
   B-frames.   It was meant to be an interrum solution for cable TV until
   MPEG-2 chips became available.

   MPEG++ is/was proprietary only at the transport layer (compatible syntax
   at the video layer).  This name was coined by the Sarnoff/Philips/
   RCA/Thomson HDTV consortium.  

   Both MPEG 1.5 and MPEG++ are now moot since MPEG-2 Simple profile and
   MPEG-2 Systems layer fill these potentials, respectively.

Q. What about MPEG-2 audio?
A. MPEG-2 audio attempts to maintain as much compatibility with        
   MPEG-1 audio syntax as possible, while adding discrete surround-sound
   channels to the orignal MPEG-1 limit of 2 channels (Left, Right or
   matrix center and difference).  The main channels (Left, Right) in
   MPEG-2 audio will remain backwards compatible, whereas new coding
   methods and syntax will be used for the surround channels.

   A total of 5.1 channels are included that consist of the two main
   channels (L,R), two side/rear, center, and a 100 Hz special effects
   channel (hence the ".1" in "5.1").

   At this time, non-backwards compatible (NBC) schemes are being
   considered as an ammedment to the MPEG-2 audio standard. One
   such popular system is Dolby AC-3.

   [installment 3: detail on Layers, AC-3, etc., optimal bitrates.]

Q. What about MPEG-2 systems?
A. [to be filled out in installment 3]
        Transport stream
        Program stream
        Timing Recovery

Q. How many bitstreams can MPEG-2 systems represent?
A. [installment 3]

Q. What are the typical MPEG-2 bitrates and picture quality?
[examples of typical frame sizes in bits]

                                        Picture type
                        I               P               B          Average
@ 1.15 Mbit/sec         150,000         50,000          20,000      38,000

MPEG-2 601              400,000         200,000         80,000     130,000
@ 4.00 Mbit/sec

Note: parameters assume Test Model for encoding, I frame distance of 15
(N = 15), and a P frame distance of 3 (M = 3).

Of course with scene changes and more advanced encoder models found
in any real-world implementation, these numbers can be very different.

Q. At what bitrates is MPEG-2 video optimal?
A. The Test subgroup has defined a few examples:

"Sweet spot" sampling dimensions and bit rates for MPEG-2:

Dimensions      Coded rate      Comments
-------------   ----------      -------------------------------------------
352x480x24 Hz   2 Mbit/sec      Half horizontal 601.  Looks almost NTSC
(progressive)                   broadcast quality, and is a good (better)
                                substitute for VHS.  Intended for film src.

544x480x30 Hz   4 Mbit/sec      PAL broadcast quality (nearly full capture
(interlaced)                    of 5.4 MHz luminance carrier).  Also
                                4:3 image dimensions windowed within 720
                                sample/line 16:9 aspect ratio via pan&scan.

704x480x30 Hz   6 Mbit/sec      Full CCIR 601 sampling dimensions.

[these numbers subject to change at whim of MPEG Test subgroup]

Q. How does MPEG video really compare to TV, VHS, laserdisc ?
A. VHS picture quality can be acheived for source film video at about
   1 million bits per second (with proprietary encoding methods).  It is
   very difficult to objectively compare  MPEG to VHS.  The response curve
   of VHS places -3 dB at around 2 MHz of analog luminance bandwidth
   (equivalent to 200 samples/line). VHS chroma is considerably less dense
   in the horizontal direction than MPEG source video (compare 80 samples/
   line to 176!).  From a sampling density perspective, VHS is superior only
   in the vertical direction (480 lines compared to 240)... but when taking
   into account interfield magnetic tape crosstalk and the TV monitor Kell
   factor, not by all that much.  VHS is prone to timing errors (which can be
   improved with time base correctors), whereas digital video is fully
   discretized. Pre-recorded VHS is typically recorded at very high
   duplication speeds (5 to 15 times real time playback), which leads to
   further shortfalls for the format that has been with us since 1977.

   Broadcast NTSC quality can be approximated at about 3 Mbit/sec, and PAL
   quality at about 4 Mbit/sec.  Of course, sports sequences with complex
   spatial-temporal activity need more like 5 and 6 Mbit/sec, respectively.

   Laserdisc is a tough one to compare.  Disc is composite video (NTSC
   or PAL) with up to 425 TVL (or 567 samples/line) response.  Thus it
   could be said laserdisc has 567 x 480 x 30 Hz "resolution". The
   carrier-to-noise ratio is typically better than 48 dB.  Timing is
   excellent. Yet some of the clean characteristics of laserdisc can be
   acheived at 1.15 Mbit/sec (SIF rates), especially for those areas of
   medium detail (low spatial activity) in the presense of uniform motion.
   This is why some people say MPEG-1 video at 1.15 Mbit/sec looks almost
   as good as Laserdisc or Super VHS.

   Regardless of the above figures, those clever proprietary encoding
   algorithms can push these bitrates even lower.

Q. Why film does so well with MPEG ?
A. Several reasons, really:

   1) The frame rate is 24 Hz (instead of 30 Hz) which is a savings of
      some 20%.  
   2) the film source video is inherently progressive.  Hence no fussy
      interlaced spectral frequencies.
   3) the pre-digital source was severly oversampled (compare 352 x 240
      SIF to 35 milimeter film at, say, 3000 x 2000 samples).  This can
      result in a very high quality signal, whereas most video cameras do
      not oversample, especially in the vertical direction.
   4) Finally, the spatial and temporal modulation transfer function (MTF)
      characteristics (motion blur, etc) of film are more ameniable to
      the transform and quantization methods of MPEG.

Q. What is the best compression ratio for MPEG ?
A. The MPEG sweet spot is about 1.2 bits/pel Intra and .35 bits/pel inter.
   Experimentation has shown that intra frame coding with the familiar
   DCT-Quantization-Entropy hybrid algorithm acheives optimal performance
   at about an average of 1.2 bits/sample or about 6:1 compression ratio.  
   Below this point, artifacts become noticable.

Q. What are some pre-processing enhancements ?

  Adaptive de-interlacing:
  This method maps interlaced video from a higher sampling rate (e.g
  720 x 480) into a lower rate, progressive format (352 x 240).   The
  most basic algorithm measures the variance between two fields, and if
  the variance is small enough, uses an average of both fields to form a
  frame macroblock.  Otherwise, a field area from one field (of the same
  parity) is selected.  More clever algorithms are much more complex
  than this, and may involve median filtering, and multirate/
  multidimensional tools.

  Pre-anti-aliasing and Pre-blockiness reduction:
  A common method in still image coding is to pre-smooth the image
  before compression encoding.  For example, if pre-analysis of a
  frame indicates that serious artifacts will arise if the picture
  were to be coded in the current condition, a pre-anti-aliasing
  filter can be applied.  This can be as simple as having a smoothing
  severity proportional to the image activity.  The pre-filter can be
  global (same smoothing factor for whole image) or locally adaptive.
  More complex methods will use multirate/multidimensional tools again.

  The basic idea of multidimensional/multirate pre-processing is to
  apply source video whose resolution (sampling density) is greater
  than the target source and reconstruction sample rates. This follows
  the basic principles of oversampling, as found in A/D converters.

  Most detail is contained in the lower harmonics anyway.  Sharp-cut off
  filters are not widely practiced, so the "320 x 480 potential" of VHS
  is never truly realized.

Q. Why use these "advanced" pre-filtering techniques?

A. Think of the DCT and quantizer as an A/D convertor.  Think of the
   pre-filter as the required anti-alias prefilter found before every
   A/D.  The big difference of course is that the DCT quantizer assigns
   a varying number of bits per sample (transform coefficient).

   Judging on the normalized activity measured in the pre-analysis
   stage of video encoding, and the target buffer size status, you have
   a fairly good idea of how many bits can be spared for the target
   macroblock, for instance.

   Other pre-filtering techniques mostly take into account: texture
   patterns, masking, edges, and motion activity.  Many additional
   advanced techniques can be applied at different immediate layers
   of video encoding (picture, slice, macroblock, block, etc.).

Q. What are some advanced encoding methods?

  Quantizer feedback
  [Thomson patent: installment 3]

  Horizontal variance [installment 3]

  motion vector cost:  this is true for any syntax elements, really.  
  Signalling a macroblock quantization factor or a large motion vector
  differential can cost more than making up the difference with extra
  quantized DFD (prediction error) bits.   The optimum can be found
  with, for example, a Lagrangian process.  In summary, any compression
  system with side information, there is a optimum point between signalling
  overhead (e.g. prediction) and prediction error.

  Liberal Interpretations of the Forward DCT        
  Borrowing from the concept that the DCT is simply a filter bank, a
  technique that seems to be gaining popularity is basis vector shaping.  
  Usually this is combined with the quantization stage since the two are
  tied closely together in a rate-distortion sense. The idea is to use
  the basis vector shaping as a cheap alternative to pre-filtering by
  combining the more diserable data adaptive properties of pre-filtering/
  pre-processing into the transformation process... yet still reconstruct
  a picture in the decoder using the standard IDCT that looks reasonably
  like the source. Some more clever schemes will apply windowing.  
  [Warning: watch out for eigenimage/basis vector orthoganality. ]

  Frequency-domain enhancements:
  Enhancements are applied after the DCT (and possibly quantization)
  stage to the transform coefficients.  This borrows from the concept:  
  if you don't like the (quantized) transformed results, simply reshape
  them into something you do like.

  Temporal spreading of quantization error:
  This method is similar to the orignal intent behind color subcarrier
  phase alternation by field in the NTSC analog TV standard: for stationary
  areas, noise does not hang" in one location, but dances about the image
  over time to give a more uniform effect.  Distribution makes it more
  difficult for the eye to "catch on" to trouble spots (due to the latent
  temporal response curve of human vision). Simple encoder models tend
  to do this naturally but will not solve all situations.

  Look-ahead and adaptive frame cycle structures:
        Scene changes
        [installment 3]        

  It is easy to spot encoders that do not employ any advanced
  encoding techniques:  reconstruced video usally contains
  ringing around edges, color bleeding, and lots of noise.


 (non-linear) Interpolation methods (Wu-Gersho)
 Convex hull projections
 Some ICASSP '93 papers, etc.

 Conformance vs. post-processing:   Post-processing makes judging
 decoder output for conformace testing near impossible.
 [installment 3]

Q. Why bother to research compressed video when there is a standard?        
A. Despite the worldwide standard, many areas remain open for
   research:  advanced encoding and pre-processing, motion estimation,
   macroblock decision models, rate control and buffer management, etc.  
   There's practically no end to it.

Q. Is so-and-so really MPEG compliant ?

A. At the very least, there are two areas of conformance/compliance in
   MPEG:  1. Compliant bitstreams  2. compliant decoders.  Technically
   speaking, video bitstreams consisting entirely of I-frames (such as
   those generated by Xing software) are syntactically compliant with
   the MPEG specification.  The I-frame sequence is simply a subset of
   the full syntax.  Compliant bitstreams must obey the range limits
   (e.g. motion vectors limited to +/-128, frame sizes, frame rates, etc.)
   and syntax rules (e.g. all slices must commence and terminate with a
   non-skipped macroblock, no gaps between slices, etc.).

   Decoders, however, cannot escape true comformance. For example, a  
   decoder that cannot decode P or B frames are *not* legal MPEG.  
   Likewise, full arithmetic precision must be obeyed before any
   decoder can be called "MPEG compliant."   The IDCT, inverse quantizer,
   and motion compensated predictior must meet the specification
   requirements... which are fairly rigid (e.g. no more than 1 least
   significant bit of error between reference and test decoders).  
   Real-time conformance is more complicated to measure than arithmetic
   precision, but it is reasonable to expect that decoders that skip
   frames on reasonable bitstreams are not likely to be considered

Q. What are some journals on related MPEG topics ?

  IEEE Multimedia [first edition Spring 1994]
  IEEE Transactions on Consumer Electronics
  IEEE Transactions on Broadcasting
  IEEE Transactions on Circuits and Systems for Video Technology
  Advanced Electronic Imaging
  Electronic Engineering Times (EE Times)
  IEEE Int'l Conference on Acoustics, Speech, and Signal Processing (ICASSP)
  International Broadcasting Convention (IBC)
  Society of Motion Pictrures and Television Engineers (SMPTE)
  SPIE conference on Visual Comminications and Image Processing
END ---------------------- CUT HERE --------------------- 2/6