CMPEG V1.0
MPEG Encoder for PC
by Stefan Eckart
June 1993
CMPEG converts a series of images into an MPEG sequence. MPEG is a
compressed (typically by a factor of 30 to 60) standardized storage
format for video. The files compressed by CMPEG can be played by
several publicly available MPEG decoders (see Appendix A).
1. Features
===========
- input formats (detected automatically):
.PPM (PBMPLUS),
.TGA (Targa 16/24 Bit)
.RAW (Image Alchemy Raw Format)
- image size up to about 512x320 (with 640 KB RAM)
- generates 'real' MPEG sequences containing I,P and B frames
- can also produce files playable by Xing software decoders
- runs under DOS / 640KB / 8086 (no '386 or Windows required)
- many coding parameters under user control
- loadable quantization matrices
- choice between several motion estimation algorithms and matching
criteria
This program is copyrighted (C) Stefan Eckart, 1993. You can use, copy
and distribute this program at no charge but you may not modify or sell
it. I don't take any responsibility regarding its fitness, usefulness
etc. (#include <your_favourite_disclaimer>). Comments, bug reports,
questions to:
Stefan Eckart
Kagerstr. 4
D-81669 Muenchen
Germany
email: stefan@lis.e-technik.tu-muenchen.de
Any feedback is welcome.
2. Introduction
===============
CMPEG converts a sequence of images into MPEG format. MPEG is a standard
format for compressed storage of video (and audio), similar in concept
to JPEG for individual images. It, however, provides higher compression
by exploiting the similarity of consecutive images of a sequence, so
called interframe coding. As with JPEG, the compression factor depends
on the source material and the desired quality of the coded sequence. A
compression to less than 5% of the original material is not uncommon. It
is thus a convenient method to keep the heavy storage requirements of
ray-traced or digitized sequences under control. The compressed
sequences can later be played back by either decompressing them in
real-time or by decoding them to an uncompressed intermediate format.
Real-time decoding is only possible for rather small image sizes,
otherwise the frame rate becomes too low. Some players/decoders are
mentioned in appendix A.
3. Usage
========
cmpeg [options] [intra.mat] [inter.mat] ctlfile framelist output.mpg
intra.mat
a textfile defining the quantization matrix for
intraframe-coded data (optional)
inter.mat
a textfile defining the quantization matrix for
interframe-coded data (optional)
ctlfile
a textfile specifying coding parameters
framelist
either a textfile containing a list of image filenames or
a template string in C 'printf' format (e.g. sample%d.tga)
output.mpg the MPEG file to be generated
Options:
-vn
-dn
-m0
-m1
-tn
-x
determines textual output produced during encoding
n=0: none (default)
n=1: summaries only
n=2: detailed information (voluminous and cryptic, for debugging)
selects cost function to be minimized by block-matching
n=0: sum of difference magnitudes (Manhattan distance, default)
n=1: sum of squared differences (Euclidian distance)
selects logarithmic block-matching
selects full-search block-matching
framelist is template starting with number n
generate Xing compatible sequences (use i.ctl as ctlfile!)
-fn
set picture rate code:
n pictures per second
-----------------------1 23.976
2 24
3 25
4 29.97
5 30
6 50
7 59.94
8 60
Examples:
To convert the Targa files sample0.tga, sample1.tga, ... sample7.tga
into an MPEG sequence sample.mpg run CMPEG with the following command
line:
cmpeg -v1 -t0 ipb.ctl sample%d.tga sample.mpg
The -v1 option enables display of the current frame number, the
number of bits for the encoded frame and the time required for encoding
the frame. -t0 informs cmpeg that it should interpret the framelist
parameter (sample%d.tga) as a template string and that it should start
with frame 0, proceeding through all successively numbered frames found.
The %d in sample%d.tga is the (C stdlib) notation for a decimal integer.
ipb.ctl is one of the control files coming with CMPEG.
You could also have created a file sample.lst explicitly listing the
names of the files. This is more flexible but requires some typing
effort. If you want to create a forward/backward loop, for example, you
would use the following sample.lst:
sample0.tga
sample1.tga
sample2.tga
sample3.tga
sample4.tga
sample5.tga
sample6.tga
sample7.tga
sample6.tga
sample5.tga
sample4.tga
sample3.tga
sample2.tga
sample1.tga
and run cmpeg with modified options:
cmpeg -v1 ipb.ctl sample.lst sample.mpg
To increase the compression factor somewhat you can switch to full
search block matching (-m1) and mean square error criterion (-d1):
cmpeg -v1 -m1 -d1 ipb.ctl sample.lst sample.mpg
This usually doesn't pay because the gain is often only in the order
of 1 or 2% and the encoding time is considerably higher.
You can also combine options:
cmpeg -v1m1d1 ipb.ctl sample.lst sample.mpg
To create a Xing compatible file from the same Targa files, which now
must be of size 160x120, call CMPEG with these options:
cmpeg -v1 -x i.ctl sample.lst sample_x.mpg
Please note that the sample files included in the distribution
(sample0.tga ... sample7.tga) are not very typical MPEG material
because they can also be efficiently compressed by conventional
archivers and the advantages of MPEG might not become too obvious from
these examples. I just didn't want to blow up the archive by including
a more realistic sequence.
4. File Formats
===============
CMPEG comes with three example control files: ipb.ctl, pvrg.ctl and
i.ctl.
You can simply use these files without modification. ipb.ctl is for real
MPEG files, whereas i.ctl has to be used to create Xing-compatible files
in conjunction with the -x option. pvgr.ctl is similar to ipb.ctl and
imitates the default coding parameters of the PVRG MPEG encoder (see
Appendix A). Use this file if you want to compare results.
The following description of the ctlfile format assumes some familiarity
with MPEG. You only need to read it if you want to create your own
control files.
4.1 Ctlfile
===========
MPEG leaves many degrees of freedom of how to encode a sequence. Some of
these choices are built into CMPEG (e.g. the decision criteria between
different block types), some are selectable by options (algorithms and
cost functions for determining the motion vectors) and some (those which
may differ from frame to frame) are defined through ctlfile.
MPEG divides a sequence into groups of pictures, containing an arbitrary
amount of pictures (but at least one). Each line in the ctlfile
corresponds to exactly one frame in each group of pictures. The first
character of this line determines the frame type: I, P or B. I frames,
or intracoded frames, are self contained, they don't depend on the
content of other frames. P frames, or predicted frames, depend on the
previous I or P frame, only the difference between the predicted and the
actual content is stored. B frames, or bidirectionally predicted frames
or interpolated frames, depend on both a previous and a subsequent
frame.
There are some rules, restricting the choice of frame type:
- the first non-B frame must be of type I
- the last frame must be of type I or P
There are also two additional rules imposed by the implementation:
- no more than eight consecutive B frames
- not more than 32 frames in total per group of pictures
The list is repeated as often as necessary to code all frames. Thus the
list can be much shorter than the sequence itself. The simplest list is
just one I frame, which produces a sequence of I frames only. This is
the mode required for Xing compatible files. Leading B frame entries are
skipped for the first group of pictures, because there is no previous I
or P frame available for interpolation. The lines are in display order,
not in decoding order.
The example ctlfile ipb.mpg defines the following pattern:
bbibbpbbpbbp
The generated MPEG stream will look like (in decoding order):
ipbbpbbpbb ibbpbbpbbpbb ibbpbbpbbpbb ...
or in display order:
ibbpbbpbbp bbibbpbbpbbp bbibbpbbpbbp ...
Remark: Decoding order is the order in which the frames are transmitted
which differs from display order in that B frames are
transmitted after the frames on which they depend.
The type character is followed by one to four parameters, depending on
the type:
I quant
P quant forward_r_size full_pel_forward
B quant forward_r_size full_pel_forward backward_r_size full_pel_backw.
quant determines how fine or coarse the image information is to be
quantized. The larger quant is, the larger are the quantization steps
and coding noise increases. The range of allowed values is 1 to 31.
The other parameters determine the range and accuracy of the motion
vectors which represent the relative position from which a particular
block of the current frame is predicted. forward_r_size and
backward_r_size are integers between 0 and 6 (inclusive), restricting
the length of the motion vectors according to the following table:
r_size
range
------------------0
-7.5..+7.5
1
-15.5..15.5
2
-31.5..31.5
3
-63.5..63.5
4
-127.5..127.5
5
-255.5..255.5
6
-511.5..511.5
If the full_pel flags are set to one, the ranges double and half pel
motions are excluded. This speeds up decoding somewhat at the expense of
a higher bit-rate.
4.2 Intra.mat and Inter.mat
===========================
MPEG allows the use of quantization matrices different from the default
ones defined in the standard. In this manner the quantization noise can
be tailored to the particular application. CMPEG can read non-default
matrices from the files intra.mat and inter.mat. Both files have to
contain a list of 64 integers each in the range 1 to 255 (inclusive).
The integers must be separated by white space (space, tab, newline). If
not specified, the matrices default to:
intra.mat:
8
16
19
22
22
26
26
27
16
16
22
22
26
27
27
29
19
22
26
26
27
29
29
35
22
24
27
27
29
32
34
38
26
27
29
29
32
35
38
46
27
29
34
34
35
40
46
56
29
34
34
37
40
48
56
69
34
37
38
40
48
58
69
83
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
inter.mat:
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
inter.mat can be specified only if intra.mat is specified also.
It is neither necessary nor recommended to include the default matrices
explicitly, it merely increases the size of the MPEG file by 64 or 128
bytes.
4.3 Infile
==========
This textfile lists the names of the images to be coded in display
order. Each line should contain exactly one filename. All specified
image files have to be of the same type and size.
In conjunction with the -t parameter, infile is interpreted as a
template string in C library syntax. This template string has to contain
exactly one % character at the position of the frame number field in the
filenames followed by one or more figures and a 'd'. Here is a list of
examples:
infile
frames
-----------------------------------------------------------f%d.tga
-> f0.tga,
f1.tga,
... f9.tga,
f10.tga, ...
f.%02d
-> f.00,
f.01,
f.09,
f.10,,
...
f%03d.tga -> f000.tga, f001.tga, ... f009.tga, f010.tga, ...
The 03 in the last example stands for: field width 3, zero padded from
the left. You could also use octal or hexadecimal numbering (o and x
instead of d).
4.4 Image Files
===============
MPEG works best for 24 bit TrueColor images. CMPEG uses Targa (Targa and
.TGA are trademarks of Truevision Inc.) format files (usually identified
by a .TGA suffix). Both top-down and bottom-up scan order is supported.
Run-length encoded files are not supported. The files can have a
resolution of 24 bit or 15 bit per pixel. Keep in mind that a 24 bit
representation of a sequence usually yields a smaller MPEG file than a
15 bit representation of the same sequence. You should use 15 bit only
when your source material doesn't have a higher resolution or if you run
out of space otherwise (24 bit Targa files are 50% larger than 15 bit
files).
CMPEG also reads the Portable Bitmap Format of Jef Poskanzers PBMPLUS
package (see Appendix A). Only the raw PPM format with 256 levels is
supported. Another format the encoder accepts is the Raw format used by
the shareware program Image Alchemy (see Appendix A) as intermediate
format. These formats were chosen because PBMPLUS and Image Alchemy
cover most image formats you will ever have as input and it is generally
faster to convert into the format used by the conversion programs
internally than into a different format, like Targa.
The file type and size is determined automatically from the first file
listed in infile. All files of a sequence have to be of the same type
and size.
If your source material is in a different format (e.g TIFF, JPEG, GIF)
you have to convert it first to one of the above formats by using
conversion programs like PBMPLUS or Image Alchemy.
5. Technical Information
========================
A major design goal for CMPEG was to keep it small enough to permit
coding of reasonable sized files on a 640 KB '286 PC (AT) without a need
for intermediate storage files. During encoding of a bidirectionally
predicted image, two reference images have to be kept in memory, which
limits maximum image size to about 512x320. This includes CIF format
(352x288).
A Xing compatible mode was implemented which creates files that can be
played on Xing MPEG players. These players are extremely fast but they
are restricted to a limited subset of valid MPEG streams. Since I don't
have information what exactly the requirements are for Xing
compatibility, the program just imitates some but not all of the
peculiarities found in files produced with Xing encoders:
- I-frames only
- only one group_of_pictures header (doesn't seem to be required but
reduces filesize)
- some additional optional fields (user data / extension data) which are
required by the decoder, probably for alignment purposes only
- temporal_reference starting with 1 instead of 0
(this is not imitated because it is a violation of the MPEG standard
and is not required by the decoder anyway)
The program was compiled into 8086 code. This makes it a bit slower than
it could be on a '386/'486. It is however still nearly twice as fast as
the PVRG code compiled with GNU GCC.
5. Known Bugs
=============
- variable bitrate streams only (no buffer feedback control)
- constrained_parameters_flag is always set irrespective of the actual
coding parameters
6. Further Reading
==================
1. Coding of moving pictures and associated audio for digital storage
media up to about 1,5 Mbit/s, Draft International Standard ISO/IEC
DIS 11172, 1992.
2. Frequently Asked Questions (FAQ) of the alt.binaries.pictures
newsgroup: contains an introduction to MPEG.
3. Documentation of the PVRG MPEG software: a thorough overview
covering many aspects of MPEG
Appendix A: Related Software
============================
This list is probably incomplete, but it's all I'm aware of. Of course
there are programs for other systems as well (Mac, Amiga etc.).
dmpeg*.zip
Offline MPEG decoder, written by me. Available from
ftp.rahul.net:/pub/bryanw/mpeg/...
mpeg_play
MPEG Video Software Decoder (Version 2.0; Jan 27, 1993)
Authors: Lawrence A. Rowe, Ketan Patel, and Brian Smith
Computer Science Division-EECS, Univ. of Calif. at
Berkeley
toe.cs.berkeley.edu:/pub/multimedia/mpeg/mpeg-2.0.tar.Z
mpegwin
Online port of mpeg_play for MS-Windows
by: Michael Simmons, msimmons@ecel.uwa.edu.au
toe.cs.berkeley.edu:/pub/multimedia/mpeg/Ports/mpegw*
(HiColor & TrueColor support, Shareware)
mplay.exe,
mpeg.exe
DOS MPEG players from Xing Technologies
(very high speed, but decodes only a small subset of the
MPEG standard)
MPEGv1.1/1.2alpha
MPEG Software Encoder/Decoder
Authors: Portable Video Research Group (PVRG)
havefun.stanford.edu:/pub/mpeg/MPEGv*.tar.Z
mpgcodec
PVRG encoder/decoder for PC compiled with GNU gcc
('386 required)
posted in alt.binaries.pictures.utilities
PBMPLUS
Portable Bit Map file conversion utilities
Author: Jef Poskanzer
garbo.uwasa.fi:/pc/graphics/pbmplus.zoo
Image Alchemy
Shareware file conversion program
Author: Handmade Software Inc.
wuarchive.wustl.edu:/msdos/graphics/alch*
-Stefan Eckart, stefan@lis.e-technik.tu-muenchen.de, June 1993.