Danny Amor's Online Library

home *** CD-ROM | disk | FTP | other *** search

/ Danny Amor's Online Library / Danny Amor's Online Library - Volume 1.iso / html / faqs / faq / mpeg-faq / part2 < prev next >

Wrap

Text File | 1995-07-25 | 47.7 KB | 1,015 lines

Subject: MPEG-FAQ: multimedia compression [2/6] Newsgroups: comp.graphics,comp.graphics.animation,comp.compression,comp.multimedia,alt.binaries.multimedia,alt.binaries.pictures.utilities,alt.binaries.pictures,alt.binaries.pictures.d,alt.answers,comp.answers,news.answers From: phade@cs.tu-berlin.de (Frank Gadegast) Date: 22 Aug 1994 12:28:48 GMT 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 reference. 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 GCA/ASCII ? 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 planned. 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. HDTV 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) specification. 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 indicated. 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 video. 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. DCT_coefficient_first 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-2. 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 routines. 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 frame. 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 ATM PES 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 MPEG-1 SIF @ 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. (interlaced) [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. Post-processing (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 compliant. Q. What are some journals on related MPEG topics ? A. 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