Path: utzoo!utgpu!watmath!att!dptg!rutgers!cs.utexas.edu!uunet!seismo!esosun!cogen!celerity!celit!billd
From: billd@fps.com (Bill Davidson)
Newsgroups: comp.graphics
Subject: Re: How to map 24-bit RGB to best EGA/VGA palette?
Keywords: RGB EGA VGA color
Message-ID: <593@celit.com>
Date: 29 Aug 89 18:35:31 GMT
References: <3129@cbnewsm.ATT.COM> <7743@cbmvax.UUCP> <13319@well.UUCP> <586@celit.com> <4862@eos.UUCP> <13381@well.UUCP> <126@vsserv.scri.fsu.edu>
Sender: daemon@fps.com
Reply-To: billd@fps.com (Bill Davidson)
Distribution: na
Organization: FPS Computing Inc., San Diego CA
Lines: 58

In article <13381@well.UUCP> gors@well.UUCP (Michael Sierchio) writes:
>Actually, doing R, G & B separately has a few problems -- that's "taxicab"
>distance, rather than Euclidean distance, for one -- the other is that such
>an error measure relies on a uniform color metric, and the RGB cube is not
>uniform, in that equidistant colors in the color cube are not equidistant
>to the "standard" observer's perceptual apparatus.

and then:
In article <126@vsserv.scri.fsu.edu> pepke@loligo.UUCP (Eric Pepke) writes:
>One could try other color spaces, such as HSV for example.  I have not done
>this.  However, I have done the brighness quantizations in perceptual space,
>that is, brightness corrected for logarithmic eyeball response and the gamma
>of the tube.  I did not spend a lot of time on this, but the few pictures
>that I tried indicated that the difference in quality was not worth writing
>home about.

Well guess what?  Somebody has tried this.  From my files:

In article <8490@venera.isi.edu> raveling@venera.isi.edu (Paul Raveling) sez:
	[this is from May in response to someone suggesting HSV]
	Maybe, but it's not clear.  About a year ago I tried some
	experiments using a different perceptual space suggested in
	a paper by Werner Frei, but got essentially the same end results
	as using RGB space.  It's possible that limitations in the
	quantization algorithm could mask the benefits of using a
	different space; the best solution probably requires coordinated
	work to evolve both the color domain and algorithms using it.

It's not clear that this is HSV but it was in response to someone who
asked about it (I don't have the paper.  Can you elaborate Paul?).
This makes sense because most of the quantizations that you'll ever do
involve colors which are going to be close in any color system.  Colors
which are close in RGB are also going to be close in HSV, HLS, YIQ, CMY,
CIEXYZ, whatever.  The relative distances will only change slightly and
relatively rarely will they cross over.  Remember that the boxes are
going to be small almost all the time.  Also, if you apply Floyd-Steinberg
dithering the box choices only have to be "close" to their optimal value.
In terms of matching a color to it's appropriate box, that's another
matter but intuitively it doesn't seem to me that it's going to make that
much of a difference.  The RGB cube is not so skewed as to be irrelevant.
Colors which are close in the cube will also be close perceptually.
If you have 24-bit's then using some other model will probably improve
your image significantly.  If you have 20000 colors and only 8 bits
to show them with then you're going to have a crude approximation no
matter what you do.  Quantization is going to eat the subtleties.

Is Roy Hall on the net?  If you're out there, what do you think of
this subject?  (I noticed that he barely touched the subject in his
book and I was wondering why).

An experiment I would *LOVE TO TRY* except I don't have any 24-bit
frame buffers to play with is to compare the various quantization
algorithms against the original 24-bit image.  I'd also like to try
and HSV scheme (and perhaps CIEXYZ) even though I don't especially
have much faith in a change for better or worse (might make a good
paper :-).

--Bill Davidson