[Theora-dev] Re: [ogg-dev] OggYUV

Timothy B. Terriberry tterribe at email.unc.edu
Tue Nov 8 10:17:46 PST 2005

John Koleszar wrote:
> I can't speak for what Theora can support today, but the VP3 source it
> derived from supported UYVY, YVYU, YUY2, and RGB24/32 source data as well.

Of course, it did this by providing its own conversion routines (which,
particularly for the RGB spaces, make assumptions about the source color
space that may NOT be true). Such routines can and should be independent
of the actual codec itself, as you suggest for the decoding (or
"extraction") side, but MS's VfW API model did not encourage this.

Chapter 4 of the Theora specification does a reasonable job of laying
out all of the possible parameters for a Y'CbCr-style color space, which
includes as a subset those needed for RGB. Much more detailed
information is available from Charles Poynton's Color and Gamma FAQs:
If you wish to do any serious video work, you should at a minimum
undestand these.

Note that Theora takes the "small number of useful formats" approach,
and can get away with it because it is _lossy_. It is expected that the
encoder will convert to the closest format Theora acutally supports, and
any information loss introduced in the process is usually trivial
compared to the quantization that will occur later. JPEG, for example,
takes a similar stance (though it supports many more formats than Theora
does). I'll note here also that YUV4MPEG actually supports much more
than 4:2:0. theora-exp's encoder_example has conversions from everything
(non-interlaced) that mjpegtools did support back when I wrote it,
including correcting the chroma field offsets. It's possible they've
added more formats since.

For a lossless codec, the luxury of a "small number of useful formats"
may not be advisable. I can't tell you how many times I've had some raw
data and been completely unable to play it with e.g., mplayer, because
mplayer did not have an apporpriate fourcc. And mplayer has made up many
of their own non-standard fourcc's (which not even all of mplayer
support) to cover for the gaping holes left after counting illi's
supposed "90% of cases on one hand". It is a common but deadly mistake
to assume that what is important to you is what is important to everyone
else. Creating a video format system around the fourcc model has always
struck me as a very, very bad idea.

Go take a look at the H.264 specification (one of the later draft
versions should be sufficient and still publicly available) for a much
broader view of the possible pixel formats and color spaces available,
and that's still just for a lossy format (though people are building
lossless compressors on top of H.264 now). And IIRC, even this doesn't
support some of the rarer 3:1:1, etc. pixel formats.

You'll also note the specification says nothing about packed formats.
Packed vs. planar is completely orthogonal to the rest of the issues,
and only arises when storing the raw pixel data directly. Supporting
either is relatively easy in software with an xstride/ystride for each
component, so there is hardly a reason not to (Theora doesn't because it
greatly simplifies several inner loops to be able to assume xstride=1; a
raw codec should not be as affected by such an issue). And there is
definitely a reason _to_ support them in a raw format, since switching
from one to the other is relatively difficult for hardware.

More information about the Theora-dev mailing list