Dan Margulis Applied Color Theory
LAB/RGB and RGB/RGB Conversions:
How Accurate?
Subj: [colortheory] sRGB to LAB conversions (was Re:
Adobe Lightroom Beta)
Date: Sunday, September 17, 2006 8:20:29 PM
From: Richard Wagner
On Sep 17, 2006, Stephen Marsh wrote:
I know that Bruce (Fraser) has later changed his
position on the critical need for these extra spaces,
but I find it
hard to accept that a very wide gamut space is now the
best option for
image editing. Capture and Rendering is one thing, but
image editing
is another thing, the problem as previously mentioned
is moving
between larger to smaller matrix profiles. We know that
there is no
perfect RGB editing space, each has it's good and bad
points.
Stephen,
This has been a very interesting discussion, and I also
appreciated your summary.
In going back and re-visiting Bruce Lindbloom's site
(hadn't been there in a while), I followed the link you provided to another
on the site, that discusses the problems in converting from 8-bit sRGB to
LAB (and back).
http:
//www.brucelindbloom.com/index.html?WorkingSpaceInfo.html
Unless I'm missing something, this seems to be an even
bigger issue than working in a wide-gamut color space, as one potentially
ends up losing a vast amount of color information due to quantization.
To quote Bruce Lindbloom:
"Another interesting observation from the table
relates to native Lab encoding. The established methods of integer encoding
of Lab color (Lab TIFF, ICC, Photoshop) will clip some of the Lab Gamut.
But even more devastating than that is the gross coding inefficiency (only
35%). This means that nearly two-thirds of Lab coding space is wasted on
colors that do not even exist. This may be seen here. This inefficiency
"squeezes" real colors tightly together, resulting in possible
quantization losses. So converting an image into Lab for the purposes of
applying a color correction in Photoshop can severly reduce the number of
unique colors in your image. This is discussed further here. Whether this
is a significant loss depends on the particular situation, but you should
at least be aware of it."
This is shown visually here:
http://www.brucelindbloom.com/index.html?
LabGamutDisplayHelp.html#IntegerLab
and a more complete discussion is given here:
http:
//www.brucelindbloom.com/index.html?RGB16Million.html
I'm not a color scientist, but this seems to be a
potentially bigger problem than working in 16-bit, wide-gamut color spaces
and later converting to smaller color spaces.
Thoughts?
--Rich Wagner
___________________________________________________________________________
Subj: [colortheory] sRGB to LAB conversions (was Re:
Adobe Lightroom Beta)
Date: Sunday, September 17, 2006 9:31:22 PM
From: Stephen Marsh
Richard Wagner wrote:
I'm not a color scientist,
Neither am I, which is why this stuff is
"interesting" and good to know about, so I don't let it get in
the way of making images look better. Being informed of the issues is good,
which is where Bruce is coming from.
A few years ago I went to the ColorSync Users list an
read in full, over 200 posts on the dangers/pitfalls of LAB (8 or 16 bpc)
and why it should be avoided.
Dan has covered this in the past on list and in his
recent LAB book.
Just as well Dan and his publisher and the market do
not agree, or at least are open to using such a "poor method"
after knowing the dangers.
but this seems to be a potentially bigger
problem than working in 16-bit, wide-gamut color spaces
and later
converting to smaller color spaces.
All true in theory, but for many users this matters
little.
Best,
Stephen Marsh.
___________________________________________________________________________
Subj: Re: [colortheory] sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Sunday, September 17, 2006 9:41:18 PM
From: Mike Russell
From: Richard Wagner
I'm not a color scientist, but this seems to be a
potentially bigger
problem than working in 16-bit, wide-gamut color spaces
and later
converting to smaller color spaces.
Thoughts?
It's not a problem in practice. More generally,
there are any number of problems that might seem to be potentially serious.
Such problems are irrelevant curiosities unless they can be illustrated by
specific images.
From a practitioner's standpoint, a theory's usefulness
is determined by whether it bears on the quality of a specific, real world
image. AFAIK, no one has shown a photograph that actually suffers
from Photoshop's Lab clipping, which is why it appears to be necessary to
resort to diagrams of rendered color gamuts to illustrate the
"problem".
Mike Russell
www.curvemeister.com/forum/
___________________________________________________________________________
Subj: [colortheory] Re: sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Monday, September 18, 2006 12:38:19 PM
From: Richard Wagner
On Sep 18, 2006, Mike Russell wrote:
AFAIK, no
one has shown a photograph that actually suffers from
Photoshop's Lab
clipping, which is why it appears to be necessary to
resort to diagrams of
rendered color gamuts to illustrate the
"problem".
Just to be clear, the problem is not
"clipping" but rather quantization loss. In the
sRGB-->Lab--sRGB round-trip, multiple, distinct RGB colors are being
lumped together as one color, at a ratio of about 8 to one. This
effect occurs throughout the sRGB gamut - not just at the borders. It
is in essence like using 7-bit RGB, with 129 steps per RGB channel
instead of 256, covering the same range. Important? Guess
that's for the end user to decide. Most of us scan at 12 - 14-bit,
even though we can't initially "see" the difference, even given
the fact that the image will likely end up as 8-bit. As a general
principle, I'm not crazy about the idea of throwing out large
percentage of the color information in an image before making any other
changes. As has been stated previously, there is no perfect working
space...
{
(2,186,238)^^1/3 = 129
2^^7 = 128
}
--Rich Wagner
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 19, 2006 9:48:22 AM
From: Marco Ugolini
It does sound ominous to hear that, due to
quantization, the 16+ million colors possible in 8-bit RGB are whittled
down to about 2 million or so after conversion to L*a*b*.
But let's remember that most images don't have even
close to that many unique colors in them. If you open an image (even one
with bright and well-defined colors, and good detail) in ColorThink Pro and
do a count of unique colors in it, you are likely to come up with numbers
in the range of 200,000 to 300,000. How high could the count possibly go?
500,000 perhaps? (I'm guessing wildly, here.)
Compared to that, upward of 2 million unique colors
would seem like a safe margin, unless there's something else I'm not
considering.
Regards.
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 19, 2006 9:53:51 AM
From: Dan Margulis
Rich Wagner writes,
Just to be clear, the problem is not
"clipping" but rather quantization loss.
No, there is no loss of any significance. Tests verify
that if you are going back and forth between sRGB and other colorspaces,
going to LAB returns a slightly larger variation from the original than
ColorMatch RGB does, but sRGB>LAB>sRGB gets you a file significantly
closer to the original than sRGB>Adobe RGB>sRGB does, and *way*
closer than sRGB>ProPhoto RGB>sRGB.
This effect occurs throughout the sRGB
gamut - not just at the borders. It is in essence
like using 7-bit
RGB, with 129 steps per RGB channel instead of
256, covering the
same range.
No, because there is no correspondence between the LAB
and RGB channels, the way that there is between the red of sRGB and the red
of ProPhoto. It would be fair to say that you're using 23-bit LAB instead
of 24-bit, but that's far more than you need. You're generating 256-level
RGB channels, but you're not looking at 256 levels in each LAB channel,
because you can't--they have nothing to do with RGB channel structure. The
only thing you can look at is all three together--more than a million
levels, from which you need to extract only three sets of 256.
That unique colors are temporarily missing is
irrelevant, if you have valid RGB channels after the conversion. The flip
answer would be to run a Gaussian Blur at .1 pixel once you return to RGB.
That would work--it would generate millions of new, statistically
significant, unique colors. There's no need, however--you will be
generating them yourself, anyway. *Any* manipulation of the new RGB
channels results in millions of new unique colors--for example, printing to
any output device (they all have to generate new channels corrected in some
way for luminosity) or even opening it on your screen--the monitor's
translation routine will generate millions of new, valid unique colors
before it displays the file.
While I hope that the above technical discussion is
useful, I agree with the other posters that the test is with real images.
If you feel that there is some damage occurring, don't give us a theory
based on gamut charts, ouija boards, histograms, and the like--get out some
real images and see what you can show us. I believe you'll find it rather
difficult. If you do, by chance produce an image that appears to show
damage, though, you should also repeat your procedure by converting the
file from sRGB to ProPhoto RGB rather than LAB. Any damage that you can
cause by conversions to LAB will be greatly magnified in the conversions to
ProPhoto.
Dan Margulis
___________________________________________________________________________
Subj: [colortheory] Re: sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Tuesday, September 19, 2006 2:55:41 PM
From: Richard Wagner
On Sep 19, 2006, at 8:38 AM, Marco Ugolini wrote:
It does sound ominous to hear that, due to
quantization, the 16+ million
colors possible in 8-bit RGB are whittled down to about
2 million or so
after conversion to L*a*b*.
Marco,
This was not the "16+ million colors in 8-bit
RGB" getting whittled down to 2 million or so, but rather any number
of colors in 8-bit RGB getting whittled down to a fraction of what you
start with. Read Bruce Lindbloom's piece again carefully. This
is a quantization effect. Multiple unique RGB colors are replaced
with a singe Lab color, at a RATIO of about 8:1. http:
//www.brucelindbloom.com/index.html?RGB16Million.html It's not like
you get to keep the 2 million to work with.
But let's remember that most images don't have even
close to that many
unique colors in them.
Agreed.
If you open an image (even one with bright and
well-defined colors, and good detail) in ColorThink Pro
and do a count of
unique colors in it, you are likely to come up with
numbers in the range of
200,000 to 300,000. How high could the count possibly
go? 500,000 perhaps?
(I'm guessing wildly, here.)
Here's a problem. ColorThink Pro is not very good
(or accurate) in counting "unique colors" in an image. It
apparently samples the image and then (? converts to lab) to do the
counting. Start with the sample image provided by Lindbloom - it will
not give an accurate pixel count or number of unique colors in ColorThink
Pro. Rather than a 4096 x 4096 pixel image with 16 million colors, CT
Pro gives a 500 x 500 pixel image with 250,000 total pixels and 210,551
unique colors - before doing anything! It is not a reliable tool for
this analysis.
I'm on the road, so I don't have my programming tools
with me. I'll look into this after I return home.
The problem is certainly real, in that color
information WILL be lost, because in the integer system that is used, there
are simply not enough Lab colors available to encode all of the sRGB
colors, by a ratio of about 8:1. It would be interesting to do this
exercise with other color spaces.
--Rich
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 19, 2006 6:16:08 PM
From: Jim Rich
I just wanted to add that in this thread it seems most
everyone is counting color in millions etc. In the printing business of
putting ink on paper with a press, my experience is that you are lucky to
get between 12 thousand and 18 thousand colors printing CMYK inks on a #1
sheet. So in the context of this thread you might want to take that
into account.
Jim Rich
___________________________________________________________________________
Subj: Re: [colortheory] Re: Adobe Lightroom Beta
Date: Wednesday, September 20, 2006 8:18:14 AM
From: Dan Margulis
Stephen writes,
Fair enough, my poor memory seemed to recall a lot of
scorn and derision
being placed upon 1.8 gamma spaces not too long ago.
Now that ProPhoto
RGB is everybody's darling, not much is made of this.
Your memory is quite good. Between about 1998 and 2001,
there was indeed a lot of berating of users (and of Apple) for not
realizing that 2.2 was far superior, because only it was perceptually
uniform, and that anyone who calibrated to any other standard (read: 1.8)
could not be considered to have a calibrated monitor, and if anyone who did
not have a calibrated monitor could only produce ninth-rate color.
Particularly, there were several attacks on me for suggesting that it
wouldn't make any difference which of the two a person used.
The two main justifications for BruceRGB were its 2.2
gamma, which, it was said, was far superior to that of ColorMatch RGB, and
the fact that it was smaller than Adobe RGB, which. it was said, is
obviously too wide-gamut for practical use.
Almost overnight, however, there was an Orwell-like
change. 1.8 gamma suddenly became a good thing. And Adobe RGB, instead of
being too wide for practical use, became too narrow. There was some
commentary at the time on the fact that simultaneously Kodak, which
believed both in ultra-wide RGBs and 1.8 gamma, was beginning to spread
around money to various parties for "consulting services", but
doubtless this was a coincidence.
That's all water under the bridge now, except for one
amusing note. You will recall that at that time, the same folk who were so
vehement about the necessity of 2.2 gamma were also strongly cautioning
that converting to LAB inflicted "catastrophic damage" on the
file, and that nobody should ever use LAB. Then, as now, the statement was
based not on any real images, but rather analysis of histograms and
gradients. The "test" of choice is: begin witha grayscale
gradient in Adobe RGB. Convert same to LAB. Observe holes in histogram.
Observe that if you convert instead to BruceRGB, there are no holes in
histogram. Conclusion: "Catastrophic damage", "Quantization
error", "Levels Loss", and "Irretrievable Data
Loss".
Since then, the same people have been reciting the same
lines, not being able to accept that people who transfer files *other* than
grayscale, or gradients, into LAB don't seem to lose levels, and not
realizing that the only thing that the "test" demonstrates is
that the L channel does not have a 2.2 gamma.
Interesting, though, now that their favored RGB does
not have the gamma they used to think was so essential, the same
"test" can be done with respect to the Adobe RGB to ProPhoto
conversion--with exactly the same results as Adobe RGB to LAB, the same
number of holes in the histogram.
Let that be a warning to all those who bought into this
demonstration of "levels loss" in LAB: beware of ProPhoto RGB!
Converting into it causes "Quantization error",
"Irretrievable data loss", "Catastrophic damage"!
Worse, for the reasons I stated in a post to Rich Wagner, conversions to
ultra-wide RGBs are inherently less reliable than conversions to LAB
because the channels are more closely related to the source space even
though the gamuts are about the same. So, anyone who believes that
conversion into LAB is a catastrophe must surely believe that conversion
into ProPhoto RGB is, well, a catastrophically catastrophic catastrophe.
Personally, I'd just look at the images and see whether
there was any damage, but to each his own.
Dan Margulis
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Wednesday, September 20, 2006 8:19:07 AM
From: Marco Ugolini
In a message dated 9/19/06 12:28 PM, Jim Rich wrote:
I just wanted to add that in this thread it seems most
everyone is counting
color in millions etc. In the printing business of
putting ink on paper with
a press, my experience is that you are lucky to get
between 12 thousand and
18 thousand colors printing CMYK inks on a #1 sheet. So
in the context of
this thread you might want to take that into account.
How do you reach that count?
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: [colortheory] sRGB to LAB conversions (was Re:
Adobe Lightroom Beta)
Date: Wednesday, September 20, 2006 12:10:42 PM
From: Olivier Desmaison
Also, so that we can progress with this, what *is* a
good and
accurate way to count unique colors in an image file,
in your opinion?
This is definitely the issue and probably why one can
always argue "you don't see it, don't bother".
The color count is meaningless in so-called achievable
values (whether RGB or Lab), but sensible in DeltaE variations. So even
though the value count is less it will be the DeltaE count (let aside
posterization and other computer related issues, since the initial
hypothesis is "real world" image)that will perceptually make the
converted file colors acceptable or not. No one would eventually care if 2
RGB values get clustered into a single one or if one is lost as long as
those two values were not intially perceptually differenciated.
Olivier Desmaison
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Wednesday, September 20, 2006 12:37:38 PM
From: RJay Hansen
On 9/20/06, Marco Ugolini wrote:
Besides the fact that this whole issue still sounds a
bit mysterious to me,
I would love to see a practical example of how this
quantization problem
affects an image, meaning something that I can see and
evaluate visually.
Watch out Marco. You're sounding like Dan.
RJay Hansen
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Wednesday, September 20, 2006 12:41:08 PM
From: Jim Rich
Marco,
Good question. I personally did not do the counting and
at this exact moment I don't know the specific scientific method used to
derive those numbers.
Over the last 30 years I have found this
information from a few sources. One source is by reading research from TAGA
proceedings, another is though my experiences with press vendors and
with colleagues at RIT.
Jim Rich
___________________________________________________________________________
Subj: [colortheory] Re: sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Wednesday, September 20, 2006 12:51:14 PM
From: Richard Wagner
"Marco Ugolini"wrote:
Is that something that you know from having tested it,
or simply an
impression?
From testing. As I described, if you start with
Bruce Lindbloom's 4096 x 4096 pixel image containing unique RGB values CT
Pro generates a 500x500 pixel image to work with (using what algorithm? No
idea...) and then reports fewer than the expected 500 x 500 = 250,000
unique pixels.
I wonder whether Steve Upton is reading this and would
care to comment.
I have posted a bug report.
Also, so that we can progress with this, what *is* a
good and accurate way
to count unique colors in an image file, in your
opinion?
Well, if you still have OS 9 running, Bruce Lindbloom
wrote a program called Levels 1.2 that did precisely this, and more. I have
an archived copy, but I don't currently have Classic running. There in not
much commercial software that I am aware of (a Corel app?) that can do
this.
"Apparently" doesn't sound convincing to me.
Can you offer something more
substantial to back your assertion?
No. I have no idea what algorithms Steve is using, or
why the data is invalid, but it clearly is. It could simply be due to the
image sampling that reduces the working image size - i.e., averaging error.
This could be tested by using smaller images that (hopefully) don't get
down-sampled.
Note that a quantization error will occur with any RGB
image, regardless of color space, that is converted to Lab - the question
is, what is the relative magnitude of the effect relative to the RGB color
space, and what is the significance in practical terms. Because the volume
of the sRGB gamut occupies a smaller percentage of the Lab gamut than
larger color spaces like Adobe or ProPhoto, I would expect the quantization
loss to be larger with sRGB than the others. I'll work on an
objective answer to this after I return home.
--Rich
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Wednesday, September 20, 2006 3:30:47 PM
From: Lee Clawson
on 9/20/06 11:58 AM, Rich Wagner at Richard Wagner
wrote:
Well, if you still have OS 9 running, Bruce Lindbloom
wrote a program
called Levels 1.2 that did precisely this, and more. I
have an archived
copy, but I don't currently have Classic running. There
in not much
commercial software that I am aware of (a Corel app?)
that can do this.
Rich,
I can run OS-9. Looked on Bruce's site for levels 1.2
but didn't see it. Any ideas where to download a copy ???
Also, while we're on the topic how many colors can the
average person actually see ????
Lee Clawson
2/\V/\7 Studio
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Wednesday, September 20, 2006 6:54:14 PM
From: Marco Ugolini
In a message dated Sep 20, 2006 7:34 AM, RJay Hansen
wrote:
Watch out Marco. You're sounding like Dan.
Ugh...
That's a risk I'll have to take.
Unlike Dan, I do not have any solid certainties to
defend, and I'm easily convinced when I am shown evidence, which so far
hasn't happened, since the whole discussion has remained purely
theoretical. In the old commercial's words, "where's the beef?".
And unlike Dan *I* don't keep raising the bar to fend
off contradiction or embarrassment... :-)
Marco
___________________________________________________________________________
Subj: [colortheory] Re: sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Friday, September 22, 2006 5:21:22 AM
From: Richard Wagner
OK, some preliminary results. All image
conversions were done in PS CS2. Where appropriate, conversions were made
using RelCol and BPC, with no dither. Image level counting was done
using Levels 1.2, (c)2001 Bruce J. Lindbloom, running under Classic 9.2 on
Mac OS 10.4.7. The test file was the 8-bit 4096 x 4096 pixel image
containing all unique colors provided by Bruce Lindbloom via his web site.
http://www.brucelindbloom.com/index.html?RGB16Million.html This is a
good file to start with, as it does not contain JPEG artifacts and it has
not been previously subjected to profile conversions that might bias the
results.
Original Image:
File Name: RGB16Million.tif
Image Dimensions: 4096 x 4096 = 16,777,216 pixels
Color Type: RGB
File Format: TIFF
Bits/Channel: 8
Color Space: sRGB (assigned in PS CS2)
There are 16,777,216 unique colors in this image.
==========
Convert the above image to Lab:
There are 2,186,765 unique colors in this image.
==========
Convert the Lab image back to sRGB:
There are 2,186,295 unique colors in this image.
Note that this data matches very closely what is
presented on Bruce Lindbloom's site. I'm not sure why the slight
difference - changes in the CMM? Also, as he stated, "back and
forth" trips to Lab do not result in increased image degradation.
It is primarily the first trip to Lab that results in quantization
loss. I have confirmed this. (additional data not shown.)
==========
OK, how about Adobe to Lab? Better than sRGB, as
I predicted in a previous post.
There are 3,135,822 unique colors in this image.
==========
And ProPhoto to Lab - should be better yet - and it is.
As expected, moving from a wide-gamut color space to Lab results in less
quantization loss than moving from sRGB to Lab.
There are 6,236,954 unique colors in this image.
==========
In testing some of my own scanned wildlife images, the
quantization loss was frequently around 50% - 70% when converting from
AdobeRGB to Lab. For example, a 14-bit scan of a Kirtland's Warbler began
with 18,617,959 unique colors; conversion to 8-bit resulted in 513,680
unique colors, and conversion of that image to Lab resulted in
279,760 unique colors.
Is any of this significant? In looking at the
images, it doesn't appear to be, although the same can be said for many
other operations that result in data loss. It is not clear what effect this
may have on an image if many operations follow. Avoiding the
transition to Lab avoids the data loss, although this obviously gives up
the advantages tof techniques unique to Lab.
--Rich Wagner
___________________________________________________________________________
Subj: [colortheory] Re: Counting unique colors...
Date: Friday, September 22, 2006 5:22:35 AM
From: Marco Ugolini
Something that works in Mac OS X, please...
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Friday, September 22, 2006 6:55:53 AM
From: Mike Russell
From: Ric Cohn
.....
So far the only one who I know has attempted this kind
of thing is
Dan. I'm not saying he has all the answers or is always
right, but at
least there's something to look at instead of talk
about!
my 2¢.
My two cents as well, Ric. I couldn't have said
it better. The appearance of the final image is the bedrock.
---
Mike Russell
www.curvemeister.com/forum/
___________________________________________________________________________
Subj: [colortheory] Re: sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Friday, September 22, 2006 12:32:11 PM
From: Richard Wagner
One last test before I leave - or before my wife wrings
my neck for making us late...
Start with what I like to call the "perfect
synthetic test image" DeltaE_8bit_gamma2.2.tif on http:
//brucelindbloom.com/index.html?ReferenceImages.html It has
dimensions of 3072 x 2048 = 6,291,456 pixels and contains no noise,
and is tagged with an ICC reference profile (sRGB primaries and white
point). It has 134,683 unique colors.
Again, using Levels to count unique colors.
Original (8-bit) image:
Red channel has 250 unique levels in the range [
0, 255]
Green channel has 253 unique levels in the range [
0, 255]
Blue channel has 253 unique levels in the range [
0, 255]
There are 134,683 unique colors in this image.
=======
Convert to Lab:
L* channel has 251 unique levels in the range [
0, 255]
a* channel has 162 unique levels in the range [ 58,
219]
b* channel has 203 unique levels in the range [ 17,
219]
There are 89,691 unique colors in this image.
=====
Convert back to sRGB:
Red channel has 251 unique levels in the range [
0, 255]
Green channel has 253 unique levels in the range [
0, 255]
Blue channel has 254 unique levels in the range [
0, 255]
There are 89,651 unique colors in this image.
This again shows that the data loss occurs in the first
trip to Lab. Not unexpected.
It would be interesting to test the other images shown.
--Rich
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: AdobeLightroom Beta)
Date: Friday, September 22, 2006 12:35:28 PM
From: Bob Frost
Lee,
Not surprising. As you age, your lenses (in your eyes)
yellow, and you therefore see less blue than when young. Anyone over 50 may
notice (or not notice) this. Plus anyone over 60 is likely to have some
clouding of the lenses which gives a nice Gaussian Blur effect!
Bob Frost.
___________________________________________________________________________
Subj: Re: [colortheory] sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Friday, September 22, 2006 12:36:36 PM
From: Matthew Rigdon
Andrew,
I downloaded the 16-bit challenge and saw the damage
you're talking about. However, Photoshop and Aperture tell me this file was
shot with a Digital Rebel (300D) at 200 ISO. Is this not correct?
I have a 300D as well and I find that it gets really
noisy in the shadows as little as 200 ISO. I just got a 20D in the last
week, so I haven't shot enough to see whether it has the same problem. I've
been told that ISO 100 and 200 are almost indistinguishable on the 20D (and
350D), but my experience with the 300D is that there's a very discernible
difference.
I can understand that working in 16-bit would be
preferable if an image is noisy to begin with, but how much noise? Would
you still have the same problem at ISO 100 on the 300D?
The file I'm looking at is CRW_0775.CRW. All my 300D
files start with this prefix. My 20D files start with IMG (actually _MG,
but I wrote an automator action to rename them in Image Capture).
Matthew Rigdon
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Friday, September 22, 2006 12:37:20 PM
From: Bob Frost
Dan,
All the figures you mention seem very low. Have a look
at this answer to the same question from Frank Welte at the University of
Massachusetts Medical School, and he says that most estimates are about 10
million colors. He also points out that people vary in the number of colors
they can see, and what color they perceive from the same color input to the
retina.
http:
//www.hhmi.org/cgi-bin/askascientist/highlight.pl?kw=&file=
answers%2Fstructure%2Fans_011.html
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Friday, September 22, 2006 12:37:56 PM
From: Mark Segal
OK, next time I'll ask my wife to assess what I've lost
from a trip back and forth between Lab and ProPhotoRGB - that'll settle it.
But really, this has to be complete nonsense. Apart from which, as I look
around me I can't believe there are 100 million colours to be seen. And how
do these bean counters distinguish between hue and luminosity? The same
solour can look different with each subtle shift of L. How often have you
had the experience of making really subtle adjustments to various Curves in
Photoshop, run the file on a printer as good as an Epson 4800 and tried
hard to see the difference? I think we're all getting too refined by half
relative to the tools we work with.
Mark Segal
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Friday, September 22, 2006 7:52:19 PM
From: Dan Margulis
Marco writes,
And unlike Dan *I* don't keep raising the bar to fend
off contradiction or
embarrassment... :-)
No, the bar is right where it's been for the past
several years.
I believe that what Marco is referring to is that those
who support his position, who are not able to provide any real-world
examples to suggest there's any merit in it, have adopted a standard
tactic. They hand in a bogus example that they know in advance I will not
accept, such as the Andrew image that Marco is now touting, and then feign
astonishment that I don't accept it, and tell the world that it proves that
I keep changing the rules and will never accept any image.
There have been a couple of most comical examples of
this already on this list, but I'll let the list enjoy one that took place
on the ColorSync list last year.
Marco posted some extravagant claims about the great
difference that 16-bit processing has made for him, and I politely inquired
as to whether by any chance he had a real-world color photograph that would
show this great advantage, which of course he did not. What he *did* have
to show the advantage was not acceptable, whereupon I was accused, as you
can see, of raising the bar, as follows:
"Dan, there is no pleasing you! You ask for an
example of banding, and I give it to you. No good, because it's a synthetic
image! Never mind that all it was meant to do was alert people to an
inherent weakness in the pixel structure of an 8 bit image, which per se
does NOT invalidate the whole edifice of 8-bit imaging. I give you another
example, this one from a B&W 16-bit scan. No good, because it's not
color! You keep raising the bar."
Or, in loose translation:
1) "The only thing you said was 'color photograph'
". How could I possibly know that you wouldn't accept a black and
white?? You keep changing the rules!"
2) "How am I supposed to know that 'photograph'
means you won't take a computer-generated graphic? You keep raising the
bar. It's a waste of time for me to show you a real example, because you're
just going to invent a reason to reject it. But they exist--trust me. I
know it for a fact."
Dan Margulis
___________________________________________________________________________
Subj: [colortheory] Re: sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Friday, September 22, 2006 7:55:25 PM
From: Andrew Rodney
The "stats" I got (somewhere years ago) were
12 million colors so this is a lot closer to the value I'd expect. Again,
this isn't the ability to see 12 million colors simultaneously! I may now
update my presentations and use 10 million to be more conservative in this
estimate.
The 100 million seemed pretty "out there"
FWIW.
Andrew Rodney
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Friday, September 22, 2006 7:55:57 PM
From: Lee Clawson
on 9/22/06 8:09 AM, Mark Segal at Mark Segal wrote:
OK, next time I'll ask my wife to assess what I've lost
from a trip back and
forth between Lab and ProPhotoRGB - that'll settle it.
But really, this has to
be complete nonsense. Apart from which, as I look
around me I can't believe
there are 100 million colours to be seen. And how do
these bean counters
distinguish between hue and luminosity? The same solour
can look different
with each subtle shift of L. How often have you had the
experience of making
really subtle adjustments to various Curves in
Photoshop, run the file on a
printer as good as an Epson 4800 and tried hard to see
the difference? I think
we're all getting too refined by half relative to the
tools we work with.
Mark,
My understanding is that they "see" unique
colors as any change. Not necessarily using our definitions that make a
clear separation of hue, value and saturation.
Lee Clawson
2/\V/\7 Studio
___________________________________________________________________________
Subj: [colortheory] "Pixel slaughter"? (was
Re: sRGB to LAB conversions)
Date: Friday, September 22, 2006 7:57:33 PM
From: Marco Ugolini
Hi Richard.
Again, we should not get wowed simply by the math. What
appears like "pixel slaughter" may turn out to be nothing of the
sort. Mine is a call for caution before we jump to conclusions, which at
this stage would seem premature, though this quantization process is
undeniable.
In the 16- vs. 8-bit debate the damage is clearly
visible (for a good example, see the RAW image which Andrew Rodney
mentioned on this forum a few days back: connect to his iDisk:
"thedigitaldog"; in the ensuing dialog box enter the name:
"public"; password: "public"; go to folder "16bit
challange"; the file is called "CRW_0775.CRW". Apply his
action to see the results).
Just as clear-cut as the evidence looks in that case,
it would also be helpful to see what this quantization damage looks like,
if it is actually visibly detectable.
Can anyone provide an example? Thank you.
Marco Ugolini
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Friday, September 22, 2006 7:58:37 PM
From: Marco Ugolini
In a message dated Sep 22, 2006 5:09 AM, Mark Segal
wrote:
How often have you had the experience of making really
subtle adjustments to various Curves in Photoshop, run
the
file on a printer as good as an Epson 4800 and tried
hard
to see the difference?
Good point.
I think we're all getting too refined by half relative
to the tools we work with.
Well, I see another pie-throwing contest coming. And I
think it serves no one to do that, other than indulge some people's
inflated idea of themselves.
Please, let's not descend yet again to a favorite
pastime on this forum, i.e., "circling the wagons" and
retrenching in a priori positions that can only be called premature before
the evidence is in. And that applies to both sides of an issue, or however
many sides there may be.
Let's do less posturing, and more debating instead,
with an open mind, possibly.
Thank you.
Marco Ugolini
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Friday, September 22, 2006 11:10:41 PM
From: Dan Margulis
In a message dated 9/22/06 12:37:21 PM, Bob Frost
writes:
All the figures you mention seem very low. Have a look
at this answer to
the same question from Frank Welte at the University of
Massachusetts Medical
School, and he says that most estimates are about 10
million colors.
I don't know where he's getting his information from. I
haven't been active in this area for several years, but I can tell you that
when I was, my opinion wasn't just on the high end--it *was* the high-end.
Every vendor put out figures that were much lower than mine.
We'll never know, because it's impossible to devise a
test that would show us--we don't have either output devices or measuring
devices with the necessary precision. Plus, there's a terminology problem.
"Different colors" means, to me, that not only can a person see
that two samples are not the same, but that there is some clue as to in
what *way* they are not the same.
Those who were giving figures in the 50,000 range were
extrapolating from average performance on the Munsell-Farnsworth test,
among others. Certainly the number of errors that the average human being
makes on that tests suggests that 50,000 is fair. My point was that there
are limitations as to how that is interpreted, and that in any case a
substantial number of people get perfect scores on that test and, inasmuch
as it was not technically possible then (and I doubt it is now) to
construct a more precise one, it's hard to know exactly *how* precise the
vision of the perfect scorers was.
Dan Margulis
P.S. The extreme figure on the low side, 2,500, was,
believe it or not, Kodak's gospel.
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Friday, September 22, 2006 11:11:47 PM
From: Mark Segal
Marco,
I have no wagons to circle and no a priori
positions - how could I - alot of this is so arcane that it would be truly
presumptuous. The purpose of my intervention was to try to correlate
principles with practice. That is often a sobering exercise, as it defines
the useful parameters of the discussion, which can go a long way to
achieving the laudable approach you describe here.
Cheers,
Mark Segal
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Saturday, September 23, 2006 4:19:21 PM
From: Mark Segal
Interesting - we now have a range from 100 milliion
colours (for my wife) to 2,500 (for Kodak). In my line of business, if I
ever presented a range for any variable that had a ratio of 1 to 40,000 it
would be the last contract I ever obtained. This is a clear symptom of a
gradullay maturing science characterized by alot of methodological
uncertainty. Hence the need to tread carefully before being carried away by
unproven and perhaps unprovable propositions.
Which brings me back to the relationship between
principles and observable reality. I have observed on the monitor (LaCie
321) and prints (Epson 4800) several real-world colour images processed in
16 bit mode from raw files in ProPhoto colour space, printed, converted the
same files to Lab mode, looked at them on the monitor, printed them, moved
back to RGB, looked at them on the monitor, re-printed them, compared the
prints and frankly couldn't detect any differences. Nor could I see any
instant differences on the monitor. I also showed the prints to my wife who
is supposed to be able to detect 100 million colours and after looking at
the outcomes she wondered why I was sitting here wasting time, ink and
paper. Maybe when monitor and printer gamuts really expand by multiples
(which judging from the recent rate of progress will be a gradual
evolution) such exercises will be worthwhile, but until then my experience
indicates it will be hard to sustain a systematic empirical relationship
between these procedures and results having a high probability of
repeatability.
Now turning to this issue of "quantization
errors" - to put my "what I'm not" cards on the table, I'm
not a colour scientist and not a professional mathematician, BUT I hope I
can formulate a certain amount of logic and ask some of the right
questions. So here goes. I was rather mystified by the notion that one
loses millions of colours shifting between RGB and Lab, especially because
I don't see the logic underlying this proposition and as explained above
couldn't detect it empirically, eventhough Bruce Lindbloom is supposed to
have explained it.
The reason I am mystified is that I'm not sure what's
going on "under the hood" between the various ways of portraying
colour using Curves. Lab has only two curves for portraying all colours in
that colour space, while RGB has three for any of its colour spaces. But
what does that mean? Is it not possible - indeed the case - that Lab *a*
includes ALL the hues along the green-magenta axis and Lab *b* includes all
the hues along the blue-yellow axis both within Lab space, which was
supposed to have been developed by the Commission Internatiionale de
l"Eclairage (CIE) based on observed reality of human visual
perception? I think Yes unless advised to the contrary. (BTW, how many
*unique colours* is that?)
Then one switches between colour spaces. What happens?
Presumably all the colours that fit within a gamut common to both get
included in the conversion. Presumably the colours that are in gamut for
the space one is moving FROM but out of gamut for the space one is moving
TO, are the ones that will be clipped or modified depending on how the
colour space conversion handles the rendering. And how many of those
affected colours are within our device gamuts say expanded by 30% to
account for projected technical progress over the next X years that could
be of any practical interest? If I could get a clearer understanding of
these questions the debate may begin to make sense to me.
Mark Segal
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Saturday, September 23, 2006 4:21:19 PM
From: Marco Ugolini
I think we should move beyond the shock of the numbers
to actually colorimetrically *measuring* what is being lost to the
"quantization effect".
I wish that someone with both enough patience and a
sufficiently scientific and methodic mind would sit down and try to sort
out how perceptually significant these quantization losses truly are.
Again, my suspicion so far is that they produce tiny
color differences, in the order of fractions of a Delta E 2000. That is to
say, well below the average viewer's ability to perceive them.
(Alright, sure, maybe those tetrachromat wonder-women
can see them...but this is not for them.)
Regards.
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: RE: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Saturday, September 23, 2006 4:21:42 PM
From: Howard Smith
Mike, I've been following these heated and often
fascinating discussions for quite awhile. As much as I agree that the
appearance of the final image is all that's really important, everyone
seems to be missing the point that these are more intellectual arguments
than practical recommendations. Each expert is certainly entitled to
his or her own opinion, but in the end all we really want is a good image.
Whatever method achieves that goal is the best method for that image.
It's doubtful that anyone involved in these discussions really
believes that everyone else will fall into line if their argument is just
presented with enough irrefutable logic. On the other hand, the
arguments do keep us from being complacent. Photoshop is such an
incredibly flexible program with such almost unlimited potential for
experimentation that it's good to keep an open mind and look for clues that
may advance our skill and knowledge. We should never fall into the
trap of arguing against things we know in our hearts to be unreasonable
without first looking to see what we can use from new ideas. Or from
old ideas that won't die a natural death. Even if a theory turns out
to be bogus, there may be some things in it that will prove very useful in
practice.
But then looking at it this way takes all the fun out
of it, doesn't it? Let the arguments continue to flourish! But maybe
without getting so personal....
Howard Smith
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Saturday, September 23, 2006 4:25:29 PM
From: Marco Ugolini
In a message dated 9/22/06 3:59 PM, Dan Margulis wrote:
I believe that what Marco is referring to is that those
who support his
position,
"Support my position"? I don't have one!
I'm still trying to figure this one out!
who are not able to provide any real-world examples to
suggest
there's any merit in it, have adopted a standard
tactic.
No, they are people in good faith, who are puzzled by
something that appears fraught with danger. Where to go next from there is
the question here.
They hand in a bogus example that they know in advance
I will not accept,
such as the Andrew image that Marco is now touting,
What you call "touting" I call "looking
with my own eyes" (certainly not yours). I see what I see, and you
would have to prove to me that I am having some serious vision problems in
order to convince me otherwise.
and then feign astonishment that I don't accept it,
So, you have already made up your mind about the
quantization effect? Please share your conclusions with us, instead of
impugning our good faith.
and tell the world that it proves that I keep changing
the rules and will
never accept any image.
There have been a couple of most comical examples
Ha ha, I'm laughing already...
of this already on this list, but I'll let the list
enjoy one that
took place on the ColorSync list last year.
Marco posted some extravagant claims
Extravagant, yes. You know, Marco is more than a bit
loopy...
about the great difference that 16-bit processing has
made for him,
So I am in Dan's formidable crosshairs now. You truly
have a big axe to grind. I never said anything of the sort. If you want to
present my position, quote me, don't freely paraphrase with such obvious
bias.
and I politely inquired as to whether by any chance he
had a real-world
color photograph that would show this great advantage,
which of course
he did not.
"Of course": has anyone told you how arrogant
you sound?
What he *did* have to show the advantage was not
acceptable,
whereupon I was accused, as you can see, of raising the
bar, as follows:
"Dan, there is no pleasing you! You ask for an
example of banding, and I give
it to you. No good, because it's a synthetic image!
Never mind that all it
was meant to do was alert people to an inherent
weakness in the pixel
structure of an 8 bit image, which per se does NOT
invalidate the whole
edifice of 8-bit imaging. I give you another example,
this one from a B&W
16-bit scan. No good, because it's not color! You keep
raising the bar."
Or, in loose translation:
1) "The only thing you said was 'color photograph'
". How could I possibly
know that you wouldn't accept a black and white?? You
keep changing the
rules!"
2) "How am I supposed to know that 'photograph'
means you won't take a
computer-generated graphic? You keep raising the bar.
It's a waste of time for
me to show you a real example, because you're just
going to invent a reason to
reject it. But they exist--trust me. I know it for a
fact."
In other words: all of you out there, here is how you
"debate", Margulis-style: Force others to adhere to *your* rules
(specially on your forum: like Mel Brooks said, it's good to be king), and
accept nothing that violates those rules (of which you are the judge,
remember!), no matter what else those examples may show or prove (banding,
etc.).
A closed mind is essential. A stalinist attitude a
helpful option.
And since you are the one making the rules, you are
guaranteed never to lose. Sweet.
And don't forget the most important part: put words in
the other person's mouth that he never actually uttered! Specially ones
that make him sound really stupid.
Got it?
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Saturday, September 23, 2006 4:33:00 PM
From: Marco Ugolini
In a message dated 9/22/06 10:02 AM, Andrew Rodney
wrote:
The "stats" I got (somewhere years ago) were
12 million colors so this is a
lot closer to the value I'd expect. Again, this isn't
the ability to see 12
million colors simultaneously! I may now update my
presentations and use 10
million to be more conservative in this estimate.
It could be that humans can see a *total* of 12 million
or so colors, but never at the same time. There could well be a *much lower
limit* to the number of colors that humans can see under any one individual
illuminant, and the 12 million total could be the sum total of *all* these
distinct viewing environments.
Incidentally, does this total take into consideration
the effects of chromatic adaptation, or instead only counts absolute
colors?
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Saturday, September 23, 2006 5:57:21 PM
From: Jim Rich
Marco,
I don't know if this test has been done recently, but a
few years ago, I and others did work in this area. This is a excerpt from
the article that was published in DO magazine. It can be found at http:
//www.jimrich.com/tips.shtml.
In December 2002 at a color management conference, I
setup and ran a test comparing eight-bit and sixteen-bit RGB images. For
this test I printed eight-bit and sixteen-bit images via Photoshop (with
profiles) to my Epson 5500 printer. Photoshop was set up to for an RGB
workflow. Nothing was done purposely to bias the test. Transparent and
reflection images were scanned by various brands of scanners as eight-bit
and high-bit images and then were converted into the Adobe 1998 working
space.
The images used were at risk to posterize due to the
combination of their image content and the extraordinary image processing
that was applied. A majority of the images in the test had over 30 edits
applied. A number of images were converted from RGB to LAB modes two or
three times. Some of the tone edits involved over a 25% change in both RGB
and LAB modes. At the conference, I placed 28 Epson prints on a table and
let a group of imaging experts (pre-press types and photographers) inspect
and review them. The test was not for color accuracy between prints.
Participants were given a form with yes and no response categories.
Test Results
The initial feedback from the group of experts who did
not choose to fill out the form but who took a few minutes to compare the
images was that they could not precisely see any differences between the
eight-bit or sixteen-bit images.
They all went on to say that any response they would
have would be a guess. This result was verified again with the
approximately 20 test forms that were filled out. The overall outcome
showed all participants were guessing (40% to 60% of the time and
they were wrong) at which images were eight-bit or sixteen-bits.
The Bottom Line
Let me make this clear. The position I have taken on
the eight-bit verses sixteen-bit argument is based on facts. It was the
advocates of sixteen-bit images who did not back up their argument with
hard evidence that peaked my interest to do this test. If I see that
evidence I am willing to reconsider my position.
Since I have spoken about this test, I have met some
high-bit end-users who will not believe a word of the results. My final
comments to them is that the hard evidence indicates in most cases that you
cannot see a difference between eight and sixteen-bit image prints. If you
do see a difference, it is only a guess.
So at least up until 2002 there was a lot of
evidence in this direction. Technology might have changed since then, and
if it has my guess is that it got better and not worse.
And. of course... others like Dan M did this type
of test well before I went there. And their conclusions were the
same.
Jim Rich
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Sunday, September 24, 2006 12:26:02 PM
From: Ron Kelly
On 23-Sep-06, at 3:03 AM, Marco Ugolini wrote:
"Support my position"? I don't have one!
Marco,
I would venture to say that your position seems to be
this: certain computer/mathematical tools indicate that switching certain
color spaces leads to loss of a huge number of colours.
Others have pointed out that there isn't any observable
effect when this happens: it can't be verified on a monitor or a print.
Seems to me we are looking for WMDs again.
Ron Kelly.
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Sunday, September 24, 2006 12:25:07 PM
From: Marco Ugolini
In a message dated 9/23/06 8:18 AM, Howard Smith wrote:
Mike, I1ve been following these heated and often
fascinating discussions for
quite awhile. As much as I agree that the
appearance of the final image is
all that1s really important, everyone seems to be
missing the point that
these are more intellectual arguments than practical
recommendations.
Right, they should be seen as starting points for
further debate. Often enough (and very understandably), a new issue gets
tackled before one has all the answers.
To be fair, I would say that people's concerns arise
from some kind of verifiable occurrences. In 8-bit quantization, unique
colors *do* indeed get "collapsed" into a smaller number of
unique colors as a consequence of a trip from flavors of RGB to Lab.
What I am personally trying to establish at this point
(since, like others, I'm still far from a conclusion) is (a) how truly
"unique" and perceptually distinct these colors are to begin
with, and (b) whether or not there exist practical ramifications, and
whether they ought to be viewed as either relevant or negligible.
To make one example, if a given group of "unique
colors" within a consecutive series (e.g., in ProPhoto RGB, with R, G
or B in increments of 1 from 0 to 255) gets collapsed after a trip to Lab,
and the difference of this group of colors from the next adjacent group is,
say, in the neighborhood of 0.5 Delta E 2000 or even less, for all
practical purposes this difference could be viewed as trivial, and those
formerly "unique colors" (now collapsed into one) could
legitimately be treated as just *one* color without undue fear.
That's what I would like to explore, as well as the
possible opposite scenario (the possibly more destructive one), plus its
implications in real-world work. A forum like this one (dedicated to
"color theory" of all things) should not fear debate of these
subjects, should it?
Each expert is certainly entitled to his or her own
opinion, but in the end
all we really want is a good image. Whatever method
achieves that goal is
the best method for that image.
As laudable as pragmatism is in most cases, a claim of
pragmatism can also become an excuse to hold on to outdated preconceived
notions that don't wish to be challenged. There *must* be room for theory
too, obviously enough, though theory as well may turn frivolous and get
sidetracked into esoterics and irrelevancies, needing to be subjected to
reality checks on a regular basis.
It1s doubtful that anyone involved in these discussions
really believes
that everyone else will fall into line if their
argument is just
presented with enough irrefutable logic. On the other
hand, the arguments do keep us from being complacent.
Hopefully. Good point.
Even if a theory turns out to be bogus, there may be
some things in it
that will prove very useful in practice.
Yes, I would say that even if a theory should prove
incomplete (without being utterly "bogus"), it may still
contribute essential bits of truth to the larger debate.
Regards.
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Sunday, September 24, 2006 12:31:03 PM
From: Marco Ugolini
In a message dated 9/23/06 7:40 AM, MARK SEGAL wrote:
Interesting - we now have a range from 100 milliion
colours (for my wife) to
2,500 (for Kodak).
The article that Andrew mentioned
(<http:
//www.post-gazette.com/pg/06256/721190-114.stm>) stated that only
tetrachromats can perceive something in the neighborhood of 100 million
colors. Tetrachromats are individuals who have a fourth functioning cone
(one more than the great majority of the rest of humans) in the area
between reds and greens, roughly in the orange range. (Kind of like a
Hexachrome-ish type of vision, only much more psychedelic!)
For genetic reasons, tetrachromats can only be women
(now, I ask of you, is that fair...?), and the estimate is that only two to
three percent of all women are tetrachromats (which would amount to roughly
as many as 100 million women on the whole planet).
So, it's not *all* women, far from it. The other 3+
billion females of the species are still struggling with limited ranges
very similar to ours (if that can make us dudes feel any better...).
So, we guys need not resent our wives or daughters
(unless they're tetrachromats, of course!).
Regards.
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: [colortheory] Re: sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Sunday, September 24, 2006 3:03:20 PM
From: Richard Wagner
On Sep 24, 2006, Jim Rich wrote:
In December 2002 at a color management conference, I
setup and ran a test
comparing eight-bit and sixteen-bit RGB images. For
this test I printed
eight-bit and sixteen-bit images via Photoshop (with
profiles) to my Epson
5500 printer. Photoshop was set up to for an RGB
workflow. Nothing was
done purposely to bias the test. Transparent and
reflection images were
scanned by various brands of scanners as eight-bit and
high-bit images and
then were converted into the Adobe 1998 working space.
Jim,
If this same test was done using a 16-bit workflow
starting from a profiled scanner, a wide-gamut working space like
ProPhotoRGB and a profiled, wide-gamut printer like the Epson 4800, and
comparisons were made to the same images scanned at 8-bit, converted to
sRGB and printed on the same printer, would you expect observers to be able
to detect the difference?
Thanks,
--Rich Wagner
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Sunday, September 24, 2006 3:06:26 PM
From: Bob Frost
Dan,
Just seen an FAQ from the Munsell Color Science Lab :-
I heard that HDMI for HDTV's can have billions of
colors but I also heard that our eyes can only see around 10 million
colors. Does this mean all that extra color doesnt matter or will it seem
clearer in some way?
The number of colors you can see at any given instant
is fairly limited, perhaps in the tens of thousands, but your visual system
is capable of adapting to different viewing conditions to make those tens
of thousands depend on what is around them. For example, in a very dark
scene in a movie, you might be able to distinguish many dark colors that
would all look alike in a bright scene. There are also issues about how the
color information is encoded, processed, and displayed that make those
seemingly extra colors helpful. The bottom line is, yes, the added
capabilities do indeed improve image quality.
Bob Frost.
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Sunday, September 24, 2006 3:06:46 PM
From: Mark Segal
Marco,
I have no problem with the legitimacy of getting
theories as correct as they can be and then trying to understand their
practical implications. But I'm stuck (and perhaps I'm the only one to have
this problem, so excuse me if I'm wasting time here) at one FACTUAL stage
before this part of the argument, and that is what lies behind the 5th word
in the extract below - *IF*?. How sure are we that anything really gets
collapsed? Is that a theory also, or has it been conclusively demonstrated
and is a consensus item amongst the colour science mathematicians? Once
there is an unambiguous expert answer to that question, one can then move
on to dealing with the extent and character of the collapse and its
practical impact on prints.
Mark Segal
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Sunday, September 24, 2006 3:11:27 PM
From: Mark Segal
Marco, thank goodness - and I'm so relieved you posted
these clarifications - heaven's forbid I should be married to a
tetrachromat on top of all the other usual issues - but maybe it wouldn't
be such a bad thing after all - there would be appreciation rather than a
bloodbath when the next wider-gamut printer comes in the door to replace
the 4800!
Cheers,
Mark Segal
___________________________________________________________________________
Subj: [colortheory] Re: sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Sunday, September 24, 2006 3:26:49 PM
From: Richard Wagner
On Sep 24, 2006, Howard Smith wrote:
As much as I agree that the appearance of the final
image is
all that's really important, everyone seems to be
missing the point
that these are more intellectual arguments than
practical recommendations. Each
expert is certainly entitled to his or her own opinion,
but in the end all
we really want is a good image. Whatever method
achieves that goal is the
best method for that image. It's doubtful that anyone
involved in these
discussions really believes that everyone else will
fall into line if their
argument is just presented with enough irrefutable
logic.
This is precisely the point. This
"debate" began with some stating that there was no
"quantization loss" in converting from sRGB to Lab. The first
step in objectively progressing the debate is determining whether or not
the "quantization loss" really exists - thus, the test with the
Levels application to answer that specific question.
and Marco Ugolini wrote:
I think we should move beyond the shock of the numbers
to actually
colorimetrically *measuring* what is being lost to the
"quantization
effect".
I wish that someone with both enough patience and a
sufficiently scientific
and methodic mind would sit down and try to sort out
how perceptually
significant these quantization losses truly are.
Again, this is right on target. The "quantization
effect" clearly exists - there is a loss of unique color information
because of the integer arithmetic that is used in the encoding and
transformations. What is still open to debate is whether or not this
loss has any "real world" significance. There has been no
data from either side of the debate on this issue - and I personally have
not "taken a position" as to whether the data loss is significant
or not. I'd like to see the objective analysis carried out further,
as Marco proposes. To paraphrase Howard Smith, this is an interesting
intellectual argument, far more than an issue that is likely to
revolutionize prepress workflows. This is the "Color
Theory" listserve, isn't it?
--Rich Wagner
___________________________________________________________________________
Subj: [colortheory] The "quantization
effect"
Date: Monday, September 25, 2006 1:20:30 AM
From: Marco Ugolini
In a message dated 9/23/06 3:33 PM, Ron Kelly wrote:
Marco:
I would venture to say that your position seems to be
this: certain
computer/mathematical tools indicate that switching
certain color spaces
leads to loss of a huge number of colours.
Others have pointed out that there isn't any observable
effect when
this happens; it' can't be verified on a monitor or a
print.
No, please, let's not do this again -- please...
First and foremost, this is NOT about the
8-bit-vs-16-bit foodfight that was "debated" so poorly in the
past on this forum.
It is a DIFFERENT topic, though it may sound similar to
some: I call it "the quantization effect", and we may want to
call it that for the sake of both brevity and precision. Please realize
that we are talking about something ELSE here. And I even changed the
subject header to reflect that.
My "position" -- if you want to call it such
-- is that:
1) The quantization effect remarked by Rich Wagner and
Bruce Lindbloom is undeniably taking place: unique colors do get collapsed
into a smaller number of unique colors after their trip from RGB to Lab.
It's just a verifiable fact.
-- BUT!
2) I have no way of knowing (at least quite yet)
whether this phenomenon affects the image in any perceptually meaningful
way. It may or it may not (my feeling is that it does *not*, by the way, if
you really want to know, but I am keeping that hunch on a leash until I can
either prove it or disprove it, because I am trying to be fair, instead of
flying some flag).
As you see, I am trying to keep an open mind about it,
with a dose of healthy skepticism thrown in, which doesn't hurt in this
forum, where people have felt compelled in the past to take sides instead
of using their own eyes and noggins (and that includes *everybody*, me
too...).
Seems to we we are looking for WMDs again.
Then that is a sad commentary on the poor quality of
the debate on this forum if professionals cannot discuss issues without
resorting to ad hominem and personal attacks divorced from the matter at
hand, degrading the exchange to a butt-heading contest.
Just like in Iraq, there are no WMD to be found in my
invitation to an open and full debate. I hope nobody will volunteer to be
the Bush of the situation. We are all capable of better
"intelligence" than that...
Regards.
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Monday, September 25, 2006 1:20:43 AM
From: Marco Ugolini
In a message dated 9/24/06 10:30 AM, MARK SEGAL wrote:
How sure are we that anything really gets collapsed? Is
that a theory also, or
has it been conclusively demonstrated and is a
consensus item amongst the
colour science mathematicians?
Mark,
There is no need to resort to the opinion of the
"big brains" to prove this: you can easily cause the
"quantization effect" on your own.
It is a *fact*: build a column (or columns) of 256
values in R, G and/or B (or a combination thereof), from 0 to 255 in
increments of 1. Assign to this RGB file a given RGB profile (e.g.,
ProPhoto RGB, or sRGB, or AdobeRGB), then convert to LAB (using either
Relative Colorimetric or Absolute Colorimetric), and finally go back to
your original RGB profile (using the same rendering intent as on the way
into LAB).
This simulates a round-trip: to LAB, and back to the
same profile with which you started.
Save this file as a TIFF and open it in GretagMacbeth's
free app ColorLab. Export the color value list for the file as text, which
you will then open in Excel.
In Excel, look at the values in the L, A, and B
columns. You will see that some cells carry exactly the same values as
other adjacent ones right above and/or below them.
That is the quantization effect in action.
Once there is an unambiguous expert answer to that
question, one can then
move on to dealing with the extent and character of the
collapse and its
practical impact on prints.
No need to delay. Since there is no denying the reality
of the quantization effect, and we don't need the blessing of an expert to
verify that, we can start examining the ramifications right now.
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Monday, September 25, 2006 1:20:47 AM
From: Dan Margulis
Mark Segal writes,
Which brings me back to the relationship between
principles and observable reality. I have observed on the monitor (LaCie
321) and prints (Epson 4800) several real-world colour images processed in
16 bit mode from raw files in ProPhoto colour space, printed, converted the
same files to Lab mode, looked at them on the monitor, printed them, moved
back to RGB, looked at them on the monitor, re-printed them, compared the
prints and frankly couldn't detect any differences.
No. You can't. It has clearly been demonstrated that
going from one RGB to LAB and back causes less variation than going from
one RGB to another and back--and these demonstrations have been both
mathematical and visual.
Everything Rich is talking about is non-LAB specific.
Every time you apply a curve to a file, you drastically reduce the number
of unique colors. Every time you move from sRGB to Adobe RGB you
drastically reduce the number of unique colors. Every time you move from
sRGB to ProPhoto RGB you reduce the number of unique colors about as much
as if you go to LAB.
What he is not understanding--and I can't blame him,
it's a difficult concept--is that this idea of "unique colors" is
valueless. The question is not how many colors the file currently contains,
but how many it will contain *on output*. And that is a problem that takes
care of itself automatically--and in fact you face it every time you work
on a file, whether or not you convert to LAB or any other colorspace.
Every time you apply a set of curves, or set up a layer
at an opacity other than 100%, or do almost any other kind of correction,
you throw away unique values. Every time you print a file, you are throwing
away unique values. By the time you're done with a typical file, you have
probably thrown away more unique values than you do by any conversion to
any colorspace.
It doesn't matter. All that throwing away unique values
means is that for any combination of XrYg, there will only be maybe 100
permissible B values, and similarly with any combinations of rb and gb.
However, the moment you start jiggering the channels individually--and any
method of outputting the file to any device does so--the missing colors
reappear, and they are statistically valid. As long as your channel
structure is good, you'll get the proper number of output colors.
Now turning to this issue of "quantization
errors" - to put my "what I'm
not" cards on the table, I'm not a colour
scientist and not a professional
mathematician, BUT I hope I can formulate a certain
amount of logic and ask some of the right questions.
There can be no error where there is no certainty of
what you are trying to get. The idea that any camera can differentiate 16
million colors is absurd. If you convert any file from any colorspace to
any colorspace and back, you'll get 1-level differences in channels a lot
of places. Even *that* is acceptable (otherwise, you could never make any
colorspace conversions at all). But what you are worrying about here is far
beyond that--it's 1-level differences in a single channel on the assumption
that the other two channels are completely accurate. If you convert a pixel
from sRGB to any other colorspace and observe that small a difference, it
is equally accurate to say that the sRGB file is erroneous as it is to say
that there has been an error in conversion.
Again--it has nothing to do with LAB, which is a fairly
gentle conversion. The real test of a conversion's efficacy is how well you
could restore the file to its original state afterward. The people who are
talking about this can't do such a test, because they don't even know what
"quantization" means, let alone "standard deviation".
But anybody who wants to test it can. You shouldn't test computer-generated
graphics, but because people are playing with this Lindbloom one, I ran a
quickie on it.
Convert one copy sRGB>LAB>sRGB ten times. Convert
another sRGB>ProPhoto>sRGB ten times, no dither in either case (the
fact that the file doesn't change between conversions isn't relevant; I've
tested that, too, with real images).
Now, compare the two files to the original. Here are
the results in the key readings, in order of importance:
Green: LAB standard deviation .69 levels, mean
variation .37.
ProPhoto standard deviation 1.31, mean variation .78.
Luminosity: LAB standard deviation .71, mean variation
.78.
ProPhoto standard deviation .86, mean variation 1.37.
Red: LAB standard deviation 1.61, mean variation 1.02.
ProPhoto standard deviation 3.13, mean variation 1.58.
If this were a real photograph instead of a
computer-generated graphic, the variations would be lower, but the relation
would be the same. Converting sRGB>ProPhoto does much, much more
violence to the file than sRGB>LAB. In this image, you'd be seeing
around twice the level of variation--frequently variations of four levels
in the critical green channel in ProPhoto; variations of even seven or
eight levels in the red.
The lesson: you can't convert to and from ProPhoto 10
times and hope to maintain file integrity, but you can go back and forth to
LAB (or ColorMatch RGB) all day long without harming anything. Unless
you're planning to convert in and out of ProPhoto that many times on a
single image, I wouldn't worry about it, but certainly you need to worry a
lot more about conversions to ProPhoto than you do LAB.
Dan Margulis
___________________________________________________________________________
Subj: [colortheory] Re: Is it art or is it a can of
soup?
Date: Monday, September 25, 2006 1:20:48 AM
From: Richard Wagner
Jim Rich wrote:
I don't know the exact answer to the question you
are posing.
But then, you only asked for my expectation.
My expectation would be no there is no observable
difference as I found in
my earlier testing. However, that is just a guess.
And I my view, there is only one way to tell if there
is observable
difference. And that is to do a fair and real world
that is intended to not
skew a test like I have outlined in my earlier work.
But even if you did prove one way or another that there
is or isn't a high
bit advantage, you would still have a bunch of
nincompoops and self severing
zealots trying to show off how big his muscles are (or
their technological
prowess).
Actually, Jim, I'm more interested not in the
"16-bit" question," but whether or not you feel that the
colors present within the gamut of the Epson 4800 that are "out of
gamut" for sRGB are perceptible by most people, or whether you feel
that the expanded gamut of devices like the 4800/9800 is just a bunch of
marketing hooey. Using aRGB potentially limits the gamut of these
devices. This does relate to the question of how many/what colors are
perceptible by (most) people.
I agree that solid data/testing would be the best way
to get a definitive answer to this question, but the perception of many is
that the expanded gamut is real and worthwhile. You could re-state my
original question using 8-bit throughout. The only reason I used 16-bit is
that there is some consensus (?) that 16-bit is wise when working in a
large-gamut working space like ProPhotoRGB. From your answer, you
seem to imply that you don't feel there would be a perceptible difference
in images printed from ProPhoto or sRGB on an Epson 4800 - i.e., that the
expanded gamut is inconsequential. Is this correct?
--Rich Wagner
___________________________________________________________________________
Subj: Re: [colortheory] "Pixel slaughter"?
(was Re: sRGB to LAB conversions)
Date: Monday, September 25, 2006 1:20:52 AM
From: Matthew Rigdon
On Sep 24, 2006, at 9:35 AM, Andrew Rodney wrote:
I can see the noise without having to resort to any
Photoshop
manipulations.
But the noise is coming from the camera being set to
ISO 200. I've noticed that the 300D (I have one as well) has some real
"problems" with noise in the shadow areas of an image at ISO 200
and up. You have to be very careful how you process these images to not
expose these problems.
Not that the noise is particularly bad, but we've
gotten used to buttery smooth images from these cameras at ISO 100. Plus,
the original Digic sensor doesn't compare to the Digic 2 for noise which is
what most of us are used to seeing (there's a lot of 20D/30D's and XT's out
there).
I thought the point of the original test as to prove
that 8-bit was the source of the noise, rather than showing that 8-bit
exacerbates noise that's already there. It's clear that you're 16-bit image
is a little cleaner than the 8-bit version, but they both have considerable
amounts of noise that need to be repaired. I'm looking at the bird feeder,
which on my computer looks pretty bad in both 16-bit and 8-bit.
Matthew Rigdon
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Monday, September 25, 2006 1:23:04 AM
From: Marco Ugolini
In a message dated 9/23/06 1:57 PM, Jim Rich wrote:
The Bottom Line
Let me make this clear. The position I have taken
on the eight-bit verses
sixteen-bit argument is based on facts. It was
the advocates of sixteen-bit
images who did not back up their argument with
hard evidence that peaked
my interest to do this test. If I see that
evidence I am willing to
reconsider my position.
Jim,
Thank you for the reply.
I'm afraid there's been a misunderstanding here. My
statement ("I wish that someone with both enough patience...")
referred to the issue of 8-bit quantization and loss of unique colors
caused by RGB-to-Lab conversions, as originally presented by Rich Wagner in
his message to this forum dated September 17, 2006, 9:31AM.
It was not meant to revive the 8-bit-vs-16-bit
"debate" (a.k.a. the never-ending slugfest, from which I try to
distance myself very carefully when in these precincts, due to its noxious
side-effects).
All the best.
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Monday, September 25, 2006 10:23:29 AM
From: Mike Russell
While it's true that individuals vary in their ability
to discern color, the article did not say that any tectrachromatic
individuals have been found.
Mike Russell
www.curvemeister.com/forum/
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Monday, September 25, 2006 11:22:04 AM
From: Mark Segal
Marco,
Thanks for pointing me to that demonstration,
which I should try. Almost at the same time, however, Dan provided quite a
comprehensive response to this set of issues. I'd be interested in your
feedback about that, as I am sure would others following this thread.
Mark Segal
___________________________________________________________________________
Subj: Re: [colortheory] Re: Is it art or is it a can
of soup?
Date: Monday, September 25, 2006 11:22:30 AM
From: Jim Rich
Richard,
I don't know the exact answer to that question. For one
I am sure the colors you get will depend on the content of the image.
That very basic.
As for the 4800/9800 is just a bunch of marketing
hooey. You might want to try and make a list of what is really being sold.
Things like better blacks and less bronzing are my understanding. And those
things are better.
IMO the 4800/9800 is a better color solution than its
predecessors. As for having more colors that are perceptible by (most)
people? At this point I would just be guessing.
Jim Rich
On 9/24/06 6:43 PM, "Richard Wagner" wrote:
Actually, Jim, I'm more interested not in the
"16-bit" question," but whether
or not you feel that the colors present within the
gamut of the Epson 4800 that are "out of
gamut" for sRGB are perceptible by most people, or
whether you feel that the expanded
gamut of devices like the 4800/9800 is just a bunch of
marketing hooey. Using aRGB
potentially limits the gamut of these devices.
This does relate to the question of how
many/what colors are perceptible by (most) people.
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Monday, September 25, 2006 11:28:24 AM
From: Bob Frost
Mike,
Well, it does say that one woman in Newcastle (UK) has
been found who from visual tests appears to have the extra visual ability.
It says that they are awaiting genetic tests to confirm that she is a
tetrachromat. And from these results, if confirmed, about 99 million women
in the world could have this four-color vision (like birds, amphibians, and
reptiles). We already know some people have only one or two types of cone,
so why should not some have four or even five? Evolution is an amazing
process!
Although we talk about a 'Standard Observer', there is
no such thing in reality. We are all different, in visual capabilities as
well as more obvious attributes.
Bob Frost.
___________________________________________________________________________
Subj: [colortheory] Is it art or is it a can of soup?
Date: Monday, September 25, 2006 2:24:48 PM
From: Jim Rich
The change of title is about Andy Warhol and the on
going debate about his art. If it is in fact art or just a picture of a can
of soup.
No one knows what the answer is but it will be debated
for a long time.
Rich,
I don't know the exact answer to the question you
are posing. But then, you only asked for my expectation. My expectation
would be no there is no observable difference as I found in my
earlier testing. However, that is just a guess.
And I my view, there is only one way to tell if there
is observable difference. And that is to do a fair and real world that is
intended to not skew a test like I have outlined in my earlier work.
But even if you did prove one way or another that there
is or isn't a high bit advantage, you would still have a bunch of
nincompoops and self severing zealots trying to show off how big his
muscles are (or their technological prowess).
While the evidence is pretty clear here about 8 vs 16
bits image editing in Photoshop, there are some rare exceptions. But
it is up to the end-user which workflow they choose. Both work quite
well.
I for one would like to see all of us to leave our
bit-depth religion at the door and act like down-to-earth adults and
keep on discussing other pertinent topics.
IMO, Lets just move on.
Jim Rich
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Monday, September 25, 2006 4:45:31 PM
From: Duffy Pratt
I'm curious. How could you devise an experiment
that showed that an individual could distinguish between 12,000,000
different colors at the same time? How long would it take to do the
experiment? At !/100th of a second per discrimination you are talking
about having a subject make distinctions for about 33 hours and 20 minutes.
But how would you ever go about seeing how many distinctions a person
can draw "at the same time" without narrowing the time per
distinction? I'm just having a hard time imagining how any such
experiment could be done, so I'm not at all surprised at the range in the
purported "results".
Duffy Pratt
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Monday, September 25, 2006 4:46:18 PM
From: Bob Frost
Mike,
Here's a quote from a research paper out of UC, San
Diego, by Jameson et al. in 2001
http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=
Retrieve&db=PubMed&list_uids=11495112&dopt=Abstract
"We use genetic analyses to examine an important
position in the gene sequence, and we empirically assess and compare the
color perception of individuals possessing more than three retinal
photopigment genes with those possessing fewer retinal photopigment genes.
Women with four-photopigment genotypes are found to perceive significantly
more chromatic appearances in comparison with either male or female
trichromat controls. We provide a rationale for this previously undetected
finding and discuss implications for theories of color perception and
gender differences in color behavior."
There are other reseach papers showing that there is
considerable genetic variation in the pigments in two of our three cones,
which cause variation in our 'normal' color perception, in addition to the
partial color blindness of 8% men and 1% women.
Bob Frost.
PS The lowly pigeon is a pentachromat! It has five
types of cones!!!!!
___________________________________________________________________________
Subj: [colortheory] Re: The "quantization
effect"
Date: Monday, September 25, 2006 4:50:46 PM
From: Marco Ugolini
In a message dated Sep 25, 2006 7:10 AM, MARK SEGAL
wrote:
Thanks for pointing me to that demonstration, which I
should try.
Almost at the same time, however, Dan provided quite a
comprehensive response to this set of issues.
Not to *this* set of issues. He did not.
I can see that there a dispiriting effort afoot to lump
this with the 8-bit/16-bit controversy and dismiss it that way, but I
consider that beneath the kind of open, honest and to-the-point debate that
professionals are supposed to be capable of engaging in.
I'd be interested in your feedback about that, as I am
sure
would others following this thread.
As far as I am concerned, the subject of the
"8-bit quantization effect" still remains not settled either way
so far (meaning, whether it is ultimately relevant or not).
------
Marco Ugolini
Mill Valley. CA
___________________________________________________________________________
Subj: Re: [colortheory] Re: Is it art or is it a can
of soup?
Date: Monday, September 25, 2006 4:52:08 PM
From: Richard Wagner
All I have been trying to find out is whether or not
you think that the expanded gamut of the newer Epson printers can produce
additional perceptible colors compared to the conditions that you used for
your test. In your test, you clipped the gamut of the inkjet printer
to AdobeRGB. If you were to repeat the test without doing so, many of
those colors that are outside the gamut of AdobeRGB (and certainly sRGB)
will be perceptible (provided that they exist in the original - obviously),
and people will likely see the difference. This has been the
experience of many people who work with fine art inkjet printing.To get
those colors, though, you will need to work in a color space that does not
limit the output of the device, as do sRGB and aRGB. I'm starting to
think this is a can of soup - alphabet soup - stirred well to obfuscate any
real discussion.
-Rich Wagner
___________________________________________________________________________
Subj: Re: [colortheory] Re: The "quantization
effect"
Date: Monday, September 25, 2006 5:07:11 PM
From: Mark Segal
Marco,
This discussion thread - as you yourself pointed
out - is not about the 8/16 bit business, which has been done to death,
resurrected and done to death again - maybe I'm missing a cycle of it!
And again, maybe I'm suffering from too much
staring at my computer monitor, but the issue I raised was about measuring
and identifying data losses converting between colour spaces and everything
that Dan said in his response to my post was focused on exactly that - if
we are looking at the same post. So I don't see why you say he is bundling
that issue with the 8/16 bit issue in order to dismiss it. He didn't
dismiss it at all. He said there is data loss, then data recovery, and in
the end outcome there is much less risk moving in and out of Lab than
moving in and out of ProPhoto. I have asked him to clarify the data
recovery aspect. So I think this discussion is on track, and let us wait
for Dan's reply to my follow-up query. I regard this as a professional
level dialogue.
Mark Segal
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Monday, September 25, 2006 6:24:58 PM
From: Mark Segal
Dan,
Once again many thanks for a lucid explanation of where
you are coming from on this issue. I am intrigued - because I don't
understand how it works - by the notion that " the problem takes care
of itself automatically" in the context I've isolated below. It would
really round-out the story if you could you explain how this happens.
Best regards,
Mark
___________________________________________________________________________
Subj: Re: [colortheory] "Pixel slaughter"?
(was Re: sRGB to LAB conversions)
Date: Monday, September 25, 2006 6:28:44 PM
From: Andrew Rodney
Let me be clear. There1s noise (there will be noise at
ISO 100 or even 50).
If you acquire the RAW file in ProPhoto RGB in 8-bit,
there1s MORE noise than if you acquire the file in ProPhoto RGB in 16-bit.
This can be seen in the green bird feeder and other areas of dark tonality.
Andrew Rodney
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Monday, September 25, 2006 8:20:41 PM
From: Richard Wagner
On Sep 24, 2006, at 8:43 PM, Dan Margulis wrote:
Everything Rich is talking about is non-LAB specific.
Nope. I've been trying to focus on the
quantization loss in going from any RGB color space to Lab, and most
specifically, in going from sRGB to Lab.
Every time you apply a curve to a file, you
drastically reduce the
number of unique colors.
Not true. Easily testable. But it begs the question,
Did you specifically test this? If so, how? I'd like to know
specifically how you are counting "unique colors" in images. Or
are you just making a guess?
Every time you move from sRGB to Adobe RGB you
drastically reduce
the number of unique colors.
FIrst of all, why would one convert from sRGB to
AdobeRGB? Regarding the reduction in unique colors - Probably - because
going from sRGB to Adobe involves a conversion through Lab, the PCS; i.e.,
sRGB-->Lab-->AdobeRGB.
Every time you move from sRGB to ProPhoto RGB you
reduce the number
of unique colors about as much as if you go to LAB.
Again, why would one convert from sRGB to ProPhoto?
Regarding the reduction in unique colors - Probably - because going
from sRGB to ProPhoto involves a conversion through Lab, the PCS. No
surprise.
What he is not understanding--and I can't blame him,
it's a difficult
concept--is that this idea of "unique colors"
is valueless.
Also not true. Sorry, Dan, but you just don't get it.
I understand that it is difficult for you, because you're married to
Lab and you've built a reputation on a Lab workflow, and you're not a
scientist. It's often hard to step back in such circumstances and be
objective, and to listen and try to understand what other people are
saying. I'll try to help you out. ;-) First, repeat after
me. "Counting the number of unique colors in an image can be a
useful exercise to measure of the loss of information content in an image
resulting from a given transformation." If you start out with
256 colors and you end up with two colors, guess what, you've lost
information! You can even demonstrate it to yourself by changing your
Monitor settings...
The question is not how many colors the file currently
contains,
but how many it will contain *on
output*. And that is a problem that takes care of
itself
automatically--and in fact you face it every time you
work on a
file, whether or not you convert to LAB or any other
colorspace.
Really? The number of unique colors in a file on output
is not a reliable measure of the quality of an image, nor is maximizing the
number of unique colors on output a good strategy for printing. If it
were, you could simply add noise to a file containing a few colors to
improve it, and that clearly doesn't happen. As I explained before -
but you apparently didn't understand - quantization loss is analogous to
working with fewer bits in the image - you lose color resolution compared
to the image that existed before the transformation. (JPG compression
does the same thing. You can't get the original colors back out of
each "square" in the compressed image, but in the jpg case the
quantization loss is spatially correlated, whereas with the sRGB-->Lab
conversion, it is not.) Because of the mathematics of encoding, for a 4096
x 4096 pixel image containing all unique colors, you are trying to code
about 8 unique colors in sRGB to every value available in Lab. The
smaller the source RGB gamut, the larger the effect of quantization loss,
since the source gamut then takes a smaller percentage of the Lab gamut,
and there is more encoding loss. This is a mathematical fact.
And as far as the problem "taking care of itself, " sure,
if you mean that you should simply be satisfied with whatever output you
end up with, that approach works. Very intellectually rewarding, too.
Let me guess - Take it on faith? The number of unique colors in
an image is simply a tool to allow one to follow, step-wise, changes that
occur in an image as a result of a specific operation or operations.
It is a marker, not an endpoint.
Every time you apply a set of curves, or set up a layer
at an opacity other
than 100%, or do almost any other kind of correction,
you throw
away uniquevalues.
Actually, many operations may **add** unique colors.
Adding contrast can increase the number of unique colors, as can adding
noise, to name but two. Most operations will *change* the values in an
image, but they don't necessarily result in a *loss* of unique colors.
Every time you print a file, you are throwing away
unique values.
Not necessarily.
By the time you're done with a typical file, you have
probably thrown away more
unique values than you do by any conversion to any
colorspace.
This completely misses the point of quantization loss,
which apparently you don't understand or are trying hard to avoid
discussing.
Again--it has nothing to do with LAB, which is a fairly
gentle
conversion.
What??? This whole conversion was about conversion to
Lab, and the loss of data that occurs when making this transformation. As
far as the quantization loss is concerned, it can easily be demonstrated
that the smaller the source gamut, the larger the quantization loss when
converting to Lab. But to get back to your last statement,
mathematically, in what way is this conversion to Lab
"gentle?" And compared to what? If specifically
discussing the conversion of RGB color spaces to Lab, ProPhoto converted to
Lab is "more gentle" than sRGB converted to Lab, as fewer
original colors in the image are lost - i.e., there is less "data
lumping." There is less disparity and a better "fit"
between the ProPhoto gamut and the the Lab gamut than between sRGB and Lab.
Most of the encoding WITHIN Lab will be out-of-gamut for sRGB, because of
the disparity in the sizes of the respective gamuts. But, heck, maybe
I don't know what "more gentle" means.
The real test of a conversion's efficacy is how well
you could restore the file
to its original state afterward. The people who are
talking about this can't
do such a test, because they don't even know what
"quantization" means, let
alone "standard deviation".
Boy, nothing like taking a condescending position to
"prove" your point. No wonder people tire of "arguing"
with you. So far, you have demonstrated that you don't comprehend
"quantization loss" - but I'm sure that fact is difficult for you
to comprehend. Instead, you simply attempt to impugn the credibility
of anyone who differs with you, and create a pile of diversionary
information for critics to muddle through.
But anybody who wants to test it can. You shouldn't
test computer-
generated graphics...
Really? Geez, someone should have told my PhD
thesis committee... All that time I wasted testing synthetic data to
check algorithms and analysis software. What a waste. And to
think that they **made** me do it. I could have finished in half the
time... all I had to do was to skip the testing and go right to the
"real data." And as a bonus, if there was an error in my
analysis or software, they probably would have never noticed!
Convert one copy sRGB>LAB>sRGB ten times. Convert
another
sRGB>ProPhoto>sRGB ten times, no dither in either
case (the fact
that the file doesn't change between conversions isn't
relevant;
If the file doesn't change between conversions, it IS
relevant. If the file is the same as the original after you have
multiple rounds of round-trip conversions, the conversion is lossless (like
LZW compression). If the file changes after every conversion, the
conversion is lossy (like JPG conversion). If the file is different
after the first conversion, but remains the same after subsequent rounds of
conversion, the overall process is lossy, but all the data loss occurs on
the first conversion, and not thereafter. Averaging 10, or
100, or one million lossless conversions (or conversions where all of the
loss is on the first round) is meaningless. There are lies, damn
lies, and (misused) statistics.
Why the sRGB-->ProPhoto--sRGB conversions? Who
cares? What does this have to do with anything, and why would anyone
do this? The parallel comparison should be ProPhoto-->Lab-->ProPhoto,
if you insist on a comparison.
What software are you using for your "statistical
analysis" and *exactly* what are you doing, so that I can attempt to
understand if there is any meaning to these gyrations, or if the intent is
simply to dazzle the impressionable. If I can reproduce your results,
perhaps I can post some meaningful comparisons.
Boy, this is not a good use of my vacation time... but
it's so hard to let misinformation float loose and so tempting on a
"color theory" listserve to force people to back up their
assertions with fundamentally sound argument. Maybe I'm just a fool... but
I've been trying to seriously understand the implications of the
RGB-->Lab quantization loss - when and where it occurs, how big the
effect is, and whether or not it is likely to have an impact on most images
in most current workflows. Sheesh!
--Rich Wagner
___________________________________________________________________________
Subj: Re: [colortheory] Re: The "quantization
effect"
Date: Tuesday, September 26, 2006 11:50:39 AM
From: Lee Clawson
on 9/25/06 5:00 PM, Mark Segal wrote:
...[snip]...He said there is data loss, then data
recovery, and in the end
outcome there is much less risk moving in and out of
Lab than moving in and
out of ProPhoto. I have asked him to clarify the data
recovery aspect. So I
think this discussion is on track, and let us wait for
Dan's reply to my
follow-up query.
I'm also confused about this data recovery. We loose
something. Then it's recovered. But the location of the hidden data is
never revealed. Is this a hidden feature of Photoshop, that is, it
computes the amount lost, looks at the image content and recreates some of
the lost data ???
This discussion thread - as you yourself pointed out -
is not about the 8/16
bit business, which has been done to death, resurrected
and done to death
again - maybe I'm missing a cycle of it!
You've got it. Without input from Adobe and some
insight to the inner workings of their RGB to LAB to RGB cycle I have
doubts that we'll understand the latest stuff. I expect the 8/16 debate to
reoccur. And take a new tack as we see and/or take more interest in high
dynamic range.
I regard this as a professional level dialogue.
I think we could do better with figuring out how to
test the ideas we're discussing.
Lee Clawson
2/\V/\7 Studio
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 26, 2006 11:53:21 AM
From: Lee Clawson
Here's a list from Google for definitions of
"quantization" ..........
http://www.google.com/search?hl=en&ie=
ISO-8859-1&q=define%3A+quantization
---------------- AND --------------------
A list of definitions for standard deviation :
http://www.google.com/search?hl=en&lr=&ie=
ISO-8859-1&newwindow=1&safe=off&c2
coff=1&q=define%3A+standard+deviation
Lee Clawson
2/\V/\7 Studio
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 26, 2006 11:55:16 AM
From: Marco Ugolini
Well, hold those horses...
For all the differences asmon individuals, the
similarities are statistically more significant. To say otherwise negates
any possibility of creating color that is pleasing or acceptable to more
than a bunch of "select people", a proposition that would very
obviously make commercial packaging or even book publishing an impossible
endeavor.
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 26, 2006 12:00:26 PM
From: Dan Margulis
Mark writes,
Once again many thanks for a lucid explanation of where
you are coming from on this issue. I am intrigued - because I don't
understand how it works - by the notion that " the problem takes care
of itself automatically" in the
context I've isolated below. It would really round-out
the story if you could you
explain how this happens.
I think we are getting to the point of diminishing
returns in discussing what happens technically. The purpose of going to LAB
is to *change* the file--not to return it unchanged to sRGB and start
comparing pixels. So the whole discussion is irrelevant in a way. Also,
many of the concepts here are quite difficult. A lot of math professors,
let alone the folks on this list, would get them wrong. But what's *easy*
is to test real images, move them back and forth several times, with and
without intervening corrections, and compare them to see whether there's
any discernible problem.
Every part of the imaging process is an ebb and flow of
eliminating previous color combinations and creating new ones to replace
them.
What I was referring to was that worrying about what
pre-printing conversions do is like worrying about a small stain on one's
shirt and then jumping into a pigsty and rolling about. If you intend to
output the image--and who doesn't intend to output it?--then you have
committed to a very violent conversion, the impact of which will dwarf that
of going to LAB. It will average certain of the RGB channels together to
make the output channels; it will rip parts out of them to construct a
black and other channels, if any; and everything gets dithered.
All these activities restore the number of unique
colors, while wiping out many distinctions that previously existed. It
happens any time you start averaging things, or moving channels around
independently.
If, for example, you make three copies of the sRGB
image, convert one each to ProPhoto RGB, Adobe RGB, and LAB, and bring them
back, all three will have lost unique colors, but in different ways. If you
make a composite of the three (assigning one-third weight to each), then
you'll get something that looks the same, but in fact has as many unique
colors as the original did. Will it handle any better? No.
Or, you can bring the LAB file back to sRGB, rotate it
5 degrees, and then rotate it back to the original position. That will
create many new unique colors, and in a statistically valid way. Does it
help in the slightest to what happens next? No.
"Unique colors" in the context of a print has
no meaning, because we don't have a tutti-frutti ink that prints all colors
simultaneously. Instead, the color is created by four or more different
applications of colorant. This insures enormous variation with or without
help from the input file, because the channels never line up exactly.
Offset printing is easiest to understand, so I'll talk in those terms, but
inkjet is conceptually just the same.
If the output file calls for a dot of 30c and a
neighboring dot at 60m, this cannot be printed as a uniform color, first
because the two dots are not centered with respect to one another, second
because the magenta dot is much larger than the cyan and thus cannot be
covered by it, third because the dots spread and are not of a uniform
density. If you're dealing with six channels, a single grouping of dots may
produce thousands of different colors. So the concept of a pixel in the
Photoshop file can never be duplicated in print.
Also, every printing dot is not computed from a single
pixel, but usually from at least three and sometimes as many as nine, all
averaged together. What is the difference between outputting a file that
has been converted 25 times sRGB>LAB>sRGB, and the same file without
the conversion?
The difference between the two files is that in each
channel, random pixels vary by one level, rarely two, almost never three.
The variation typically occurs only in one channel at a time, and there is
no telling whether the variation will be lighter or darker.
Now, what does it take to *see* a difference in print?
Remember, you're evaluating very small dots, each of which may be partially
covered by other dots. I believe it is inconceivable that you could detect
that one dot of one channel had varied by one level (=<.5 percent in
printing terms). I believe it is extremely unlikely that you could detect
that an individual dot had varied by two levels (=<1% in printing
terms), but, for the sake of argument, I grant that you can.
Assuming 3-4 pixels are averaged together, and assuming
that the standard deviations that Photoshop reports are correct, the odds
of such an event occurring can be computed. For a file that's been
converted back and forth 25 times, the odds of a single halftone dot
varying by two or more levels from the original are around a hundred
billion to one. If it did occur, you could never detect it, because it
would be hidden by the natural variation of the output device. SWOP
permits, in the midtone, variations of the equivalent of six or seven
levels in dots.
This is why people don't have to fear conversions--it
takes a whole lot to produce enough variation to be noticeable. As I
pointed out earlier, if you convert in and out of ProPhoto 10 times, you
may achieve that. I have shown in other tests that if you convert in and
out of Adobe RGB 25 times, you may have damaged the file enough to be
visible. But LAB? AFAIK, you can convert and reconvert indefinitely, and
the file is just as good as the original for any conceivable purpose.
Off to Canada now. I'll be offline for a week.
Dan Margulis
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 26, 2006 12:07:32 PM
From: Dan Margulis
Rich writes,
Nope. I've been trying to focus on the
quantization loss in going
from any RGB color space to Lab, and most specifically,
in going from
sRGB to Lab.
Yes, and I've been trying to explain that you're
chasing rainbows, worrying about an effect that takes place with almost any
move you might make in Photoshop.
FIrst of all, why would one convert from sRGB to
AdobeRGB?
I don't know, but I do know that in 1998 Adobe strongly
recommended this workflow for Photoshop 5. The conversion from sRGB to
Adobe RGB was described at that time by Adobe, Bruce Fraser, and others, as
"virtually lossless" and we were assured that it could not
possibly degrade the image. Since most of the people who made these
assurances and who thought that this was a good workflow at the time are
associated with many of the concepts you espouse, I thought that the fact
that sRGB>LAB is demonstrably gentler than the "virtually
lossless" sRGB>Adobe RGB is relevant.
Again, why would one convert from sRGB to ProPhoto?
Probably they bought into some of the hype of those who
claim that ultra-wide gamut RGBs are necessary if you want to take full
advantage of some new printer's gamut. Or, they did not wish to work in
ProPhoto, but they needed a ProPhoto copy because they wanted to use one of
its channels. As for ProPhoto to sRGB, this has been repeatedly advocated
on this list by several parties.
The more pertinent question seems to me, if you are so
very satisfied with your file that you are concerned about losing
microscopic variations in it, why are *you* going sRGB>LAB? The normal
reason for going to LAB is that people are *dissatisfied* with what they
currently have and wish to make major changes.
The entire question of "loss", therefore, is
ultimately meaningless. Yes, as a matter of curiosity, sRGB>LAB doesn't
lose anything, and as a matter of curiosity you can move in and out of LAB
at least 25 times in a given image and still have an image that's
equivalent to the original for any conceivable purpose.
But if these things were *not* true, as you seem to
think, who cares? The very fact that you move into LAB indicates that you
don't like the sRGB data, so it's unimportant whether you ever get back to
where you were. And, as to whether going in and out of LAB 25 times damages
anything, I leave it to the people who use that workflow to worry about it.
First, repeat after me.
"Counting the number of unique colors in an image
can be a useful
exercise to measure of the loss of information content
in an image
resulting from a given transformation."
I have worked with color for many years. You are the
very first person in that time who has suggested such a thing. There is no
reason for it to be true, no reason why this would affect reproduction, and
no reason for anyone to desire more unique colors in their images.
The theory is contradicted by the experimental results.
Going sRGB>Adobe RGB loses many fewer unique colors, yet sRGB>LAB is
significantly less damaging. Going sRGB>ProPhoto loses a comparable
number of unique colors to LAB, but is *far* more damaging.
If you start out with 256
colors and you end up with two colors, guess what,
you've lost
information!
True enough, but I'm not starting out with 256 and I'm
not ending with two. I'm starting out with 100,000, padding it out to
10,000,000 with random digits, and reducing it to 1,000,000. It is not
obvious that I have lost information.
Actually, many operations may **add** unique colors.
Adding contrast
can increase the number of unique colors, as can adding
noise, to
name but two. Most operations will *change* the values
in an image,
but they don't necessarily result in a *loss* of unique
colors.
That's right, and nobody said otherwise. Since the
point of your post was that you did not wish to throw away existing unique
values, I commented that any method of correction throws some of them away.
That others may be added, I did not construe as being relevant to your
thinking.
This completely misses the point of quantization loss,
which
apparently you don't understand or are trying hard to
avoid discussing.
I do understand it, thank you, but the concept is not
applicable to this topic. I also understand the dangers of imputing
statistical significance to random numbers, which is the whole problem with
your theory.
All that time I wasted testing synthetic data to check
algorithms and
analysis software. What a waste. And to
think that they **made** me
do it. I could have finished in half the time...
all I had to do was
to skip the testing and go right to the "real
data."
If your theory is intended to describe what happens
when you convert black and white computer-generated gradients to LAB, then
there is some merit in what you say. But few of use convert either black
and whites or computer-generated gradients to LAB. We convert color
photographs, so ultimately you have to test with those. If your testing
with computer-generated graphics suggests one result but you can't
replicate it with color photographs, then the testing is not relevant.
If the file doesn't change between conversions, it IS
relevant. If
the file is the same as the original after you have
multiple rounds
of round-trip conversions, the conversion is lossless
(like LZW
compression). If the file changes after every
conversion, the
conversion is lossy (like JPG conversion).
You are confusing compression and conversion
algorithms. They're two different animals. In colorspace conversions,
having one-level differences is normal and expected behavior, of no
statistical significance.
Maybe I'm just a fool... but I've been trying to
seriously understand the implications of the RGB-->Lab
quantization loss - when and where it occurs, how big the effect is,
and whether or not it is likely to have an impact on most images in
most
current workflows. Sheesh!
It's not foolish to try to learn something about the
process, and the question of what is going on during the conversion to LAB
is relevant. However, you are making a lot of unwarranted assumptions--such
as that this "loss" occurs at all. I would suggest that you
satisfy yourself that it doesn't by doing some back-and-forth manipulations
with real images, not histograms or gradients. Then, if you have some loss
that you can show us, I'm sure the list would be interested to see it. I've
tested this quite comprehensively, have published the results, and they are
quite decisive. But, there's always room for more tests.
Dan Margulis
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 26, 2006 12:34:51 PM
From: Dan Margulis
Lee Clawson writes,
Here's a list from Google for definitions of
"quantization"
And for those who care to read it, it indicates that
those who allege that a "quantization error" is going on during a
colorspace conversion are stating that the original file is continuous-tone
analog.
Dan Margulis
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 26, 2006 12:35:27 PM
From: Henry Davis
On Sep 25, 2006, at 7:20 PM, Richard Wagner wrote:
Boy, nothing like taking a condescending position to
"prove" your point.
Now I've heard the pot call the kettle black, again!
I believe that it is a valid concern that there is some
mystery in determining the number of unique *printed* colors. It is
not merely a statistical proposition. Along with an assertion or
claim about printable unique colors comes the question of what it is to be
a "meaningful" color - not just a mathematically unique.
I hear in this thread a desire by some to keep the
discussion within the realm of "color spaces", but
When a printer, or ink manufacturer, makes the claim of
having a larger color gamut, how is it that the claim is validated?
Is the method simply the measuring of the volume of the solids and
overprints, and thereby extrapolating the number of unique colors? If
this is the case, I have to ask if it is possible that an ink set could be
developed that would have a super large volume, yet very poor printing
would result from its use.
I seems to me that there would come a point of
diminishing returns. A point where, either mathematically or
experientially, there is nothing gained in a larger color volume.
Sorry, but the existential parts of color theory are relevant - and
this part doesn't fit into a neat equation. Again, when a printer or
ink manufacturer makes a claim, even one that says "you can see the
difference", is this a load of hooey, and how can we know? Has
the limit been reached already?
Some of this is sounding humorous, again. I can
imagine mathematicians who have never held a real ruler and never actually
measured anything because they don't have to because the math is correct,
therefore there is no need for actually measuring. If I can't see a
real, discernible difference, personally it doesn't matter a whit to me
just how many integers are lost. I am, however, concerned about
claims, and how they are justified.
Henry Davis
___________________________________________________________________________
Subj: [colortheory] sRGB to LAB conversions (was Re:
Adobe Lightroom Beta)
Date: Tuesday, September 26, 2006 6:05:35 PM
From: Andrew Rodney
On 9/26/06 10:08 AM, "Henry" wrote:
When a printer, or ink manufacturer, makes the claim of
having a larger
color gamut, how is it that the claim is validated?
Is the method
simply the measuring of the volume of the solids and
overprints, and
thereby extrapolating the number of unique colors?
Nothing to do with unique colors. If you measure the
primaries (which every package that builds an ICC profile would ask to be
printed), you can easily calculate the gamut boundaries of this color space
and produce a gamut map.
In order to fully define the unique colors, you1d have
to output 16.7 million color samples and measure them. Possible but not at
all practical (and you1d end up with an ICC profile that would be larger in
file size than most images). So when you output say 4000 color patches and
measure them, there1s a good deal of interpolation going on to build the
profile. But the pure primaries as well as paper white and pure black are
patches in any profile target. So unique colors? We can1t say but gamut,
yes.
Again, when a printer or ink manufacturer makes a
claim,
even one that says "you can see the
difference", is this a load of
hooey, and how can we know? Has the limit been
reached already?
In some cases where you can1t measure the claim, yes,
it probably is hooey. But many of the claims can be measured (and when I
say measured, I1m including just looking at the results of two variants).
Andrew Rodney
___________________________________________________________________________
Subj: [colortheory] sRGB to LAB conversions (was Re:
Adobe Lightroom Beta)
Date: Tuesday, September 26, 2006 6:05:55 PM
From: Andrew Rodney
On 9/26/06 9:46 AM, "Dan Margulis"
wrote:
I don't know, but I do know that in 1998 Adobe strongly
recommended this
workflow for Photoshop 5.
They did? Where and when, I1d like to read more about
this.
There1s only one reason I can fathom why anyone would
do this. You have a pile of sRGB images and you need to composite them with
other images that are in Adobe RGB (1998). Otherwise, there1s absolutely no
reason you1d make such conversions.
Probably they bought into some of the hype of those who
claim that ultra-wide
gamut RGBs are necessary if you want to take full
advantage of some new
printer's gamut.
Probably? I1d still like to see any text you have
archived that states why anyone would believe this. Converting sRGB into
Adobe RGB (1998) doesn1t provide any additional color gamut unless you
composite new wider gamut data or start painting something.
Or, they did not wish to work in ProPhoto, but they
needed a
ProPhoto copy because they wanted to use one of its
channels. As for ProPhoto
to sRGB, this has been repeatedly advocated on this
list by several parties.
Again, you1re guessing and the guess provides
questionable logic on the part of those so called parties.
Converting from ProPhoto RGB to sRGB makes lots of
sense. Going the other way? Nope.
The more pertinent question seems to me, if you are so
very satisfied with
your file that you are concerned about losing
microscopic variations in it,
why are *you* going sRGB>LAB? The normal reason for
going to LAB is that
people are *dissatisfied* with what they currently have
and wish to make major
changes.
Agreed. That1s the reason we alter ANY RGB or CMYK
number in a file. No one is debating that. The question is, what1s the best
way to alter the numbers and what are the ramifications.
The entire question of "loss", therefore, is
ultimately meaningless.
To you but this debate would never be discussed if
everyone agreed to this premise. Apparently lots of people are interested
in knowing the 3color theory2 of altering pixel values and the over all
effect over time to the data when output to all kinds of devices.
Yes, as a matter of curiosity, sRGB>LAB doesn't lose
anything, and as a matter
of curiosity you can move in and out of LAB at
least 25 times in a given image
and still have an image that's equivalent to the
original for any conceivable
purpose.
Equivalent?
But if these things were *not* true, as you seem to
think, who cares? The
very fact that you move into LAB indicates that you
don't like the sRGB data,
so it's unimportant whether you ever get back to
where you were.
I totally agree with one part of your statement. The
second part about what1s unimportant is a stretch to many on this list. If
that were not true, such lengthy and regular posts would never appear.
Andrew Rodney
___________________________________________________________________________
Subj: RE: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 26, 2006 6:08:47 PM
From: Randall.N.Hoffner
R.W.G. Hunt is, since 1967, Visiting Professor of
Physiological Optics
at the City University of London (i.e., he probably has
a clue what he
is talking about from a scientific standpoint).
He is also a poet, and
from the Prologue to his book, "Measuring
Colour":
"This is the story of Mister Chrome
who started out to paint his home.
The paint ran out when half way through
so to the store he quickly flew
to buy some more of matching hue,
a delicate shade of egg-shell blue.
But when he tried this latest batch
He found it simply didn't match,
No wonder he was in a fix,
for of the colors we can mix,
The major shades and those between,
ten million different can be seen."
And:
"the moral of this tale is clear:
to understand just what we see,
object, light, and eye, all three,
must colour all our thinking through
of chromic problems, old or new!"
Hunt, R.W.G., "Measuring Colour", New York,
John Wiley and Sons, 1987.
He doesn't mention anything about a fourth cone.
Randy Hoffner
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 26, 2006 6:09:04 PM
From: Richard Wagner
On Sep 26, 2006, at 8:45 AM, Dan Margulis wrote:
I think we are getting to the point of diminishing
returns in discussing what
happens technically.
Only if you step away from a rigorous analysis of what
happens with every step along the way. I see a lot of hand-waving on
this List, but seldom a rigorous analysis of anything.
The purpose of going to LAB is to *change* the
file--not
to return it unchanged to sRGB and start comparing
pixels.
The purpose of ANY operation on an image is to change
the file - that part is obvious. Ideally, you SHOULD be able to make a
"round-trip" to another color space and end up where you started.
That is the whole principle of ICC Color Management. If I created the
"RandomizedWorkingSpace" that resulted in a complete
randomization of the mapping of image pixels, it wouldn't be very useful,
would it? Or a working space where everything looked beautiful, but
converting from that space to any output space inverted the colors, or
averaged adjacent colors, or made every other pixel black, or otherwise
mangled the image. The "problem" with Lab, as we know it, is in
the integer arithmetic implementation in most current software, based on a
decision that was made years ago to gain processing speed and minimize
storage space. (Integer arithmetic is computationally faster than
floating point arithmetic.) There is nothing inherently "wrong"
with Lab, the color space; in fact, the concept was brilliant. The
"problem" is in how the pixels get represented and pushed around,
mathematically. This is described here: http:
//www.brucelindbloom.com/index.html?WorkingSpaceInfo.html
So the whole discussion is irrelevant in a way. Also,
many of the concepts here
are quite difficult. A lot of math professors,
let alone the folks on this list, would get
them wrong.
I disagree. This is an educated list, and I would
bet that anyone on this List is capable of understanding nearly any concept
if it is explained clearly. I spent 10 years doing stochastic math and
working with theories built on random processes. I could explain the
concepts to anyone here in about 15 minutes. This list understands a
lot of complex math concepts, even if they don't think about them from a
mathematical perspective. For example, take binning and histograms.
Most people on this list could describe how to take a grayscale image
and build a histogram, and explain how different images would give
different histogram "shapes." I'll bet that, given time,
they could also explain this to someone off the street, but I doubt that
most could formally describe, mathematically, what they were doing.
The big picture in life ain't that complicated.
But what's *easy* is to test real images, move them
back and forth
several times, with and without intervening
corrections, and
compare them to see whether there's any discernible
problem.
This is the antithesis of the scientific method. The
problem with this approach is that the "discernible problems" may
only show up if you look at an enormous number of images. Picking 5
or ten "representative" images may completely miss the problem.
So you go away with the impression that there is no problem, and
later it surfaces. But then, you're not looking for it, and you now
"believe" that there is no problem, so you go looking,
fruitlessly, for a different problem. Your energy gets focused in the
wrong direction, because your initial assumption is wrong. Scientists
of all disciplines use synthetic data specifically to look for problems in
their analysis. If you can run the synthetic data through a process and you
get what you expect, great (as long as you realize that the synthetic data
may not cover the real-worlld possibilities that you're interested in).
On the other hand, if the process with synthetic data does NOT give
you what you expect, STOP! You've discovered a problem and it needs
to be tracked down and addressed! What is the problem, when and where
does it occur, is it a problem with the model, the analysis, or the
synthetic data? Often these problems would never be discovered using
"real" data. You start with "simple" synthetic
data, and increase its complexity to more closely simulate real-world data
- e.g., by adding different types of noise to the data, or by modifying it
in other ways. When you can analyze and explain all of your synthetic data,
move on to "real" data. This is part of the scientific method.
You can use simple synthetic data to look at the
effects of color profile transformations and other operations. Andrew
Rodney describes how to build the Granger Rainbow in Photoshop here: http:
// www.ppmag.com/reviews/200510_rodneycm.pdf. Bruce Lindbloom built,
and provides, many synthetic images just for this purpose. Lindbloom
is a color scientist. He has the tools/brains/techniques to generate
far more complex synthetic data than most of us can. He provides
these images to the Color Community as a public service - along with
the rest of the information on his site. If you're interested in
"Color Theory" look it over, and learn to use it. It is a
very valuable resource. http://www.brucelindbloom.com/
The rest of this discussion was too far off-track from
the original question about the RGB-->Lab quantization loss for me to
pursue. I still feel that the question of the significance of this
effect has not been satisfactorily answered.
--Rich Wagner
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 26, 2006 6:18:08 PM
From: Mark Segal
Dan, there are two levels of discussion here and I
think we shouldn't blur them if we want to fully understand the issue.
There is an empirical/operational level, and there is a theoretical level;
I don't think it is strictly kosher to dismiss the latter with what
may be perfectly valid points about the former.
I like to think I'm a practical, results-oriented
person, so I respect the idea that if one does something to an image that
doesn't make a visible difference, why worry any further (the empricial
level)?
But that only takes us so far. The people who take
issue with this approach would argue that device gamuts and colour
differentiation capabilities will improve over time, hence one should seek
to degrade the files as little as necessary while achieving photographic
objectives, just in case one wants to reprint on better equipment say ten
years from now. In these circumstances, one does want to understand IN
PRINCIPLE what is happening mathematically when making these conversions.
There are, it seems, two ways of doing this: (i) understanding the
mathematics of colour space conversions, and (ii) using the colour-counting
software (if reliable for the purpose) that others in this discussion have
cited as demonstrating conclusively that millions of colours are collapsed
and destroyed converting from RGB to Lab. But that in itself is not fully
meaningful unless one can also explain in principle how much of a colour
collapse is needed to be actually visible in real-world prints.
Are there any colour scientists in this Group who could
explain the theoretics in a manner that lesser-mortals can understand?
Mark Segal
___________________________________________________________________________
Subj: RE: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Tuesday, September 26, 2006 11:02:00 PM
From: David Cardinal
and (ii) using the colour-counting software (if
reliable for
the purpose) that others in this discussion have cited
as
demonstrating conclusively that millions of colours are
collapsed and destroyed converting from RGB to Lab. But
that
in itself is not fully meaningful unless one can also
explain
in principle how much of a colour collapse is needed to
be
actually visible in real-world prints.
Are there any colour scientists in this Group who could
explain the theoretics in a manner that lesser-mortals
can understand?
No research I've ever read has provided a single
convincing answer to that question. There are too many variables and a lot
depends on whether you get to pick the colors, etc., and what you assume
about output possibilities.
Just to provide some context and to come back to Dan's
point that every typical Curves move destroys theoretical colors I did the
quick math for a simple Curve with one control point that pulls 128 for
each channel down to 108 (or up to 148). This eliminates 20 values on each
channel (which you can actually see if you have an accurate enough
histogram) and makes the theoretical number of colors 215*215*215 = 9.9M,
so we have vaporized around 6 million of our original 16M colors with quite
a trivial move. But I doubt any of us think of a move like that as making
any of our actual photographs "less colorful".
--David Cardinal
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Wednesday, September 27, 2006 2:07:17 PM
From: Marco Ugolini
In a message dated 9/24/06 8:43 PM, Dan Margulis wrote:
It has clearly been demonstrated that going from one
RGB to
LAB and back causes less variation than going from one
RGB to another and
back--and these demonstrations have been both
mathematical and visual.
Here is a homemade test devised by yours truly (a test
that our resident critic will surely find a way to pan), using my
"unscientific" tools and inferior education, a test which I
encourage other similarly unscientific and uneducated colleagues to
replicate, following these steps:
1) Make an RGB file containing 6 pixels across per 256
down (6 x 256 = 1,536 total pixels, distributed in 6 columns and 256 rows);
each pixel will be a unique color (except 5, as we will see).
2) First column: first pixel at the top is R 255, G 0,
B 0; create intermediate pixels going down in decrements of 1, all the way
to R 0, G 0, B 0 (so, R 255, 254, 253, etc, whereas G and B remain at 0 all
the way down). Total pixels in column: 256.
3) Second column: first pixel at the top is R 0, G 255,
B 0; create intermediate pixels going down in decrements of 1, all the way
to R 0, G 0, B 0 (so, G 255, 254, 253, etc, whereas R and B remain at 0 all
the way down). Total pixels in column: 256.
4) Third column: first pixel at the top is R 0, G 0, B
255; create intermediate pixels going down in decrements of 1, all the way
to R 0, G 0, B 0 (so, B 255, 254, 253, etc, whereas R and G remain at 0 all
the way down). Total pixels in column: 256.
5) Fourth column: first pixel at the top is R 255, G
255, B 0; create intermediate pixels going down in decrements of 1, all the
way to R 0, G 0, B 0 (so, both R and G 255, 254, 253, etc, whereas B
remains at 0 all the way down). Total pixels in column: 256.
6) Fifth column: first pixel at the top is R 255, G 0,
B 255; create intermediate pixels going down in decrements of 1, all the
way to R 0, G 0, B 0 (so, both R and B 255, 254, 253, etc, whereas G
remains at 0 all the way down). Total pixels in column: 256.
7) Sixth column: first pixel at the top is R 0, G 255,
B 255; create intermediate pixels going down in decrements of 1, all the
way to R 0, G 0, B 0 (so, both G and B 255, 254, 253, etc, whereas R
remains at 0 all the way down). Total pixels in column: 256.
8) This file contains 1,536 pixels, but only 1,531
unique colors (1,536 - 5 = 1,531. That's because there are 6 pixels that
read R 0, G 0, B 0, so only one out of 6 is unique).
9) Using this file, do the following profile-to-profile
conversion tests via the "Convert to Profile" command in
Photoshop using this file, saved in TIFF RGB format. In all instances, the
conversions in this test are made using the Relative Colorimetric rendering
intent with black point compensation and no dither. (Though RGB
matrix-based profiles already use relative colorimetric in all cases, this
is a procedural detail that I want to adopt knowingly, even with
conversions to Lab, which use the Perceptual and Absolute Colorimetric
intents, but not RelCol. In any case, the point is to maintain the same
settings in all conversions as a procedural norm.)
Start each conversion by making sure to assign your
source profile first (be it sRGB, AdobeRGB or ProPhoto RGB, which are the 3
I used).
The results of my conversions (using Photoshop CS2
followed by a calculation of unique colors made using the application
Levels 1.2) are as follows:
PART 1 -- RGB-to-RGB Conversions:
From AdobeRGB to ProPhoto RGB, then back to
AdobeRGB: from 1,531 to 1,437 unique colors. Unique colors lost: 6.14%
From AdobeRGB to sRGB, then back to AdobeRGB:
from 1,531 to 1,367 unique colors. Unique colors lost: 10.71%
From sRGB to AdobeRGB, then back to sRGB: from
1,531 to 1,436 unique colors. Unique colors lost: 6.21%
From ProPhoto RGB to AdobeRGB, then back to
ProPhoto RGB: from 1,531 to 1,368 unique colors. Unique colors lost: 10.65%
From sRGB to ProPhoto RGB, then back to sRGB:
from 1,531 to 1,413 unique colors. Unique colors lost: 7.71%
From ProPhoto RGB to sRGB, then back to ProPhoto
RGB: from 1,531 to 1,317 unique colors. Unique colors lost: 13.98%
PART 2 -- RGB-to-LAB-to-RGB Conversions:
From AdobeRGB to LAB, then back to AdobeRGB:
from 1,531 to 1,297 unique colors. Unique colors lost: 15.28%
From ProPhoto RGB to LAB, then back to ProPhoto
RGB: from 1,531 to 1,251 unique colors. Unique colors lost: 18.29%
From sRGB to LAB, then back to sRGB: from 1,531
to 1,290 unique colors. Unique colors lost: 15.74%
Summation of results:
RGB-to-RGB conversions lose a minimum of 6.14%
and a maximum of 13.98% of the original file's unique colors (average loss:
9.23%). (Please note that the maximum value here is derived from the
ProPhoto RGB >sRGB > ProPhoto RGB conversion, which is inherently
destructive by its own nature, due to the extreme gamut mismatch between
the two profiles.)
RGB-to-LAB-to-RGB conversions lose a minimum of
15.28% and a maximum of 18.29% of the original file's unique colors
(average loss: 16.44%).
Conclusions (the ones that come to my attention first,
though others may see more):
The minimum percent of unique colors lost in RGB
> LAB > RGB conversions (15.28%) is 2.5 times higher (249%) than the
minimum percent of unique colors lost in RGB-to-RGB conversions (6.14%).
The maximum percent of unique colors lost in RGB
> LAB > RGB conversions (18.29%) is 1.3 times higher (131%) than the
maximum percent of unique colors lost in RGB-to-RGB conversions (13.98%).
The lowest rate of loss in RGB > LAB > RGB
conversions (15.28%) is 1.15 times higher (115%) than the highest rate of
loss in RGB-to-RGB conversions (13.98%).
I have included the results of the ProPhoto RGB >
sRGB > ProPhoto RGB conversion though, I must repeat, I believe it skews
the overall results of the test somewhat unfairly: there is such a gamut
disparity between those two profiles that the number of unique colors lost
*must* be high by definition. Nonetheless, even with that result included,
there is still a markedly higher rate of loss in RGB > LAB > RGB
conversions.
In light of this fact, the statement that "it has
clearly been demonstrated that going from one RGB to LAB and back causes
less variation than going from one RGB to another and back" would
certainly appear either mistaken or a product of wishful thinking. This
test unequivocally demonstrates *the opposite*.
The question is not how many colors the file currently
contains, but
how many it will contain *on output*.
No, the question at this stage of inquiry is how many
unique colors go in and how many come out. Anything else is an attempt to
rush to conclusions and influence the results.
And that is a problem that takes care of itself
automatically--
Sounds like magic...
and in fact you face it every time you work on a file,
whether or not
you convert to LAB or any other colorspace.
Meaning...?
Every time you apply a set of curves, or set up a layer
at an opacity other
than 100%, or do almost any other kind of correction,
you throw away unique
values. Every time you print a file, you are throwing
away unique values. By
the time you're done with a typical file, you have
probably thrown away more
unique values than you do by any conversion to any
colorspace.
Why say "probably" when there are ways to
find out exactly how many unique colors get "thrown out" in these
types of conversions? Don't we wish to know what is happening in detail, if
and when possible, so that we can analyze what is occurring and later
decide how relevant this is in practice?
To be satisfied with "probably" is very
incurious.
It doesn't matter. All that throwing away unique values
means is that for any
combination of XrYg, there will only be maybe 100
permissible B values, and
similarly with any combinations of rb and gb.
Now you lost me there. Are you trying to blow smoke, or
would you care to explain what that means?
However, the moment you start jiggering the channels
individually--and any
method of outputting the file to any device does
so--the missing colors
reappear, and they are statistically valid. As long as
your channel structure
is good, you'll get the proper number of output colors.
The missing colors "reappear", as if they had
gone on vacation for a while? Based on what, if I may ask? Faith or
evidence? And who or what determines a "proper number of output
colors"?
Again--it has nothing to do with LAB, which is a fairly
gentle conversion.
Not according to the tests above. Unless it's a matter
of faith, in which case I let people believe what they will.
The real test of a conversion's efficacy is how well
you could restoret the
file to its original state afterward. The people who
are talking about this
can't do such a test, because they don't even know what
"quantization" means,
"Quantization": the "process of
approximating a continuous range of values... by a relatively-small set of
discrete symbols or integer values." In other words, we replicate an
unbroken analog signal (audio, video, etc.) by dividing it up into
components of a given size: the more subdivisions per unit of measure (the
higher the sampling rate and/or the bit depth), the more accurate the
results will be, and closer to the form of the analog signal (the Nyquist
rule being one of the most widely-used ways to determine optimal results).
In most of imaging, calculations are based on 256
levels per channel (in tristimulus color modes like RGB, LAB and XYZ),
which is a relatively small number of discrete values. Therefore a degree
of rounding and approximation is always occurring. These approximations are
cumulative through the stages of image manipulation, and their compounded
effect is larger when the sampling units are larger (more coarse).
So, though some people may not know what
"quantization" means, it's not so abstruse a concept that you
couldn't try to explain it in your own words, instead of leaving it at
"they don't even know what "quantization" means". So
tell them.
let alone "standard deviation".
Again, why, instead of throwing a generalized
accusation at your interlocutors, don't you at least *try* to explain to us
in your own words how this "standard deviation" becomes relevant
in this discussion? Unless it's a not too subtle form of ad hominem, or
smearing.
But anybody who wants to test it can. You shouldn't
test computer-generated
graphics, but because people are playing with this
Lindbloom one, I ran a
quickie on it.
Which "Lindbloom one" exactly? The 16.7
million color image, or the one called "Delta E"?
I will assume you mean the 16.7 million colors one,
which is the one that was mentioned in the email from Rich Wagner that
started this thread.
Convert one copy sRGB>LAB>sRGB ten times.
OK, though I am not sure about the relevance of the
"10 times".
After 10 sRGB>LAB>sRGB conversion cycles, the
file retains 2,106,315 unique colors out of the original 16,777,216: 87.45%
of the original number of unique colors are lost, 12.55% retained.
Convert another sRGB>ProPhoto>sRGB ten times, no
dither in either case
After 10 sRGB>ProPhoto>sRGB conversion cycles,
the file retains 2,679,215 unique colors out of the original 16,777,216:
84.03% of the original number of unique colors are lost (a figure 3.41%
*lower* than in the sRGB>LAB>sRGB conversion), and 15.97% are
retained.
The conclusion -- obvious to me -- is that
sRGB>LAB>sRGB still loses *more* unique colors than
sRGB>ProPhoto>sRGB (though, to be fair, not by a very large number).
Now, compare the two files to the original. Here are
the results in the key
readings, in order of importance:
Green: LAB standard deviation .69 levels, mean
variation .37.
ProPhoto standard deviation 1.31, mean variation .78.
Luminosity: LAB standard deviation .71, mean variation
.78.
ProPhoto standard deviation .86, mean variation 1.37.
Red: LAB standard deviation 1.61, mean variation 1.02.
ProPhoto standard deviation 3.13, mean variation 1.58.
I truly don't see (a) what these numbers are supposed
to mean and (b) how they were calculated. You give us no data to support
them.
Also, would you please explain why you are using
"standard deviation" and "mean variation" when what we
are supposed to be considering at this stage of the discussion are *unique
colors*, i.e., how many there are at the beginning and how many we end up
losing?
If this were a real photograph instead of a
computer-generated graphic, the
variations would be lower, but the relation would be
the same. Converting
sRGB>ProPhoto does much, much more violence to the
file than sRGB>LAB.
Makes no sense to me at all. The numbers I calculated
provide the exactly opposite conclusion.
In this image, you'd be seeing around twice the level
of variation--frequently
variations of four levels in the critical green channel
in ProPhoto;
variations of even seven or eight levels in the red.
What does this have to do with unique colors, how many
we have and how many we lose? You are talking past us (at least me). I am
not sure how productive that is, though I know how condescending.
The lesson: you can't convert to and from ProPhoto 10
times and hope to
maintain file integrity, but you can go back and forth
to LAB (or ColorMatch
RGB) all day long without harming anything.
"You can't convert...and hope"? Based on
"standard deviation" and "mean variation"? What logic
does a conclusion like this follow?
Unless you're planning to convert in and out of
ProPhoto that many times
on a single image, I wouldn't worry about it, but
certainly you need
to worry a lot more about conversions to ProPhoto than
you do LAB.
If you say so...
--------------
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: Re: [colortheory] Re: The "quantization
effect"
Date: Wednesday, September 27, 2006 2:14:53 PM
From: Bob Frost
Mark,
Data recovery? Surely it must be data 'invention'.
Bob Frost.
___________________________________________________________________________
Subj: Re: [colortheory] Re: Is it art or is it a can
of soup?
Date: Monday, September 25, 2006 11:22:30 AM
From: Jim Rich
Richard,
I don't know the exact answer to that question. For one
I am sure the colors you get will depend on the content of the image.
That very basic.
As for the 4800/9800 is just a bunch of marketing
hooey. You might want to try and make a list of what is really being sold.
Things like better blacks and less bronzing are my understanding. And those
things are better.
IMO the 4800/9800 is a better color solution than its
predecessors. As for having more colors that are perceptible by (most)
people? At this point I would just be guessing.
Jim Rich
___________________________________________________________________________
Subj: RE: [colortheory] Re: sRGB to LAB conversions
(was Re: AdobeLightroom Beta)
Date: Wednesday, September 27, 2006 5:38:25 PM
From: David Cardinal
Lee--Two points:
1) Per my typo correction, it is actually 40 levels
lost
2) That loss is per channel. Remember that you get to
16M as 255*255*255, so to calculate the number you have left with 215 per
channel you multiply 215*215*215 or about 9.9M.
Thus the "loss" of 6M+ of our original unique
color possibilities.
Dan makes a good point. But one that should be known.
In the
case of the round trip to LAB I had no idea there'd be
loss.
I always knew that quantization would cause loss in
that case, but personally I think the color counting doesn't actually help
answer whether it is meaningful loss or not.
--David Cardinal
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Wednesday, September 27, 2006 5:41:34 PM
From: Richard Wagner
David,
Thanks for contributing to this discussion.
I agree with you in principle, but not in detail. I'm
not sure where you get 215*215*215. I get (256-20)**3, or 13,144,256,
which can be confirmed by counting the number of unique levels in the
resulting file with the Levels app. This results in a loss of
3,632,960 unique colors from the original 16,777,216 unique colors in this
image. As you describe, the loss of levels is quite easy to see on an RGB
histogram, even in PS CS2 (try it!); there are 20 sharp "spike
canyons" on the left where the levels are lost, and 20 "spike
peaks" on the right where the resulting levels of changed pixels
overlap pre-existing levels, raising the frequency count from 1 to 2. (I
used a +20 change.)
This raises an important point. When looking at
synthetic data, it is important to understand how that data differs from
real-world data. In this case, because all possible colors are
represented in the image, changing any pixel color will, of necessity,
result in the loss of that color, and the creation of duplicates of some
other colors. Put another way, each pixel in the image will no longer have
a unique color - i.e., there will be multiple pixels with the same color,
as can be clearly seen in the right-hand half of the histogram. Thus, with
this image as a starting point, there must be a loss of unique colors with
any curve manipulation - any curve move will not be able to create new
unique colors. That is not true for an image that is not
"color-dense" - in other words, an image that has few colors can
be subjected to a curve manipulation without the loss of unique colors,
because although some of the original colors are lost in the move, new
colors can be created. This was my basis for stating that curves
moves do not of necessity result in the loss of unique colors in an image.
It depends on the image that you start out with. The "16 million
colors" image represents one extreme, and a black-and-white image
would represent the other extreme. "Real-life" color photo
results will be somewhere in between.
Perhaps more interesting than the previous question is
what happens if you compare the unique colors in the original image after
conversion to Lab, with the same image subjected to the +20 Curves
transform, then converted to Lab. The "huge loss" of unique
colors from the curves move is dwarfed by the move to Lab. The original
image drops to 2,186,765 unique colors, whereas the +20 curves move image
ends up with 2,096,928 unique colors. The numbers barely move after
conversion back to sRGB: 2,186,295 vs 2,096,651. This also gets
closer to the typical "round-trip" RGB-->Lab-->RGB workflow
that many are interested in, as most images are not distributed in Lab.
Of course, none of this answers the more important
question of "how far" the colors change from their original
values. If the changes are 1/256 levels for R, G, or B, they are not
likely to be of any real significance. If, on the other hand, the
"changed" colors were up to 10 levels away, we would obviously
all have a problem with this. I'm working on a more complete analysis
of this more important question, and the preliminary results indicate that
the magnitude of the change is small, at least from sRGB-->Lab--sRGB.
I'm working on a lot of other related questions, as well, but I'd
like to have a chance to get the analysis completed and carefully review it
myself before dumping it on the list. It's actually been an
interesting exercise using the tools at hand (primarily Photoshop), rather
than writing custom software to do the analysis, or using Mathematica or
Matlab. I'd like to ultimately present this in a manner that all can
follow (and potentially critique), and be able to replicate on their own.
Maybe we'll all learn something.
Best,
--Rich Wagner
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Wednesday, September 27, 2006 6:12:48 PM
From: Mark Segal
David,
Sorry, on thinking further I'm having two problems with
all this: (1) the item at issue is loss of colours due to conversion from
one colour space to another. It isn't clear to me how a demonstration of
what happens when altering curves is relevant to what happens when you
convert a file from colour space A to colour space B.
(2) sRGB fits into ARGB98 fits into Lab, hence as one
moves a file from one space to the next the effect should depend on what
space one starts with and in which direction one is moving. For some moves
if there is only remapping, one would expect the impact to be minimal. For
other moves, one can understand that clipping or compression of OOG colours
would occur and have a larger impact on the original file. I come back to
my earlier point that apart from the empirical demos, it would be nice to
hear what mathematical issues under the hood cause data losses exceeding
what one would expect in some of these circumstances.
Mark Segal
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Wednesday, September 27, 2006 6:16:11 PM
From: Jim Rich
Marco,
It is very interesting what you have done.
And if I understand it you have clearly shown one
side of this issue. That is there is some loss of information as you change
between modes and icc profiles.
My observation, as well as others, is that there are
two sides to this.
And I liken all of the arguments to being placed on a
scale that has a pivot point in the middle that balances the two arguments.
The science/math argument is on one side. And the end-user
experience (what they see and detect on a hardcopy proof) on the other
side. And then the scale goes back-and-fourth as people state their
cases.
You have done the math (quite well I might add) as
others have, but when a group of 8 and 16 bit images are prepared
(with heavy image editing) and laid out for experts to examine, those
experts cannot definitively tell you which print is 8 or 16 bit. And
if they are an honest bunch they will admit they are just guessing.
Then the scale swings back toward the end-user
experience that points out it doesn't really matter (in most cases) if you
use 8 or 16 bits for image editing in Photoshop. That is unless you mix
science with religion.
Does this mean that the 8 bit folks are right and that
the 16 bit folks are wrong? Or vice versa? That depends on who you
ask.
Or, does it mean that yes there is loss of data, but
even experts cannot definitively see the difference? It looks that way.
Perhaps it it art or perhaps it just a can of soup. The
debate goes on.
Jim Rich
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Wednesday, September 27, 2006 6:17:06 PM
From: Mike Russell
From: Ron Kelly
It's too bad this list is called "Color
Theory" as opposed to say,
"Dan Margulis Color Correction Discussion" or
somesuch.
Here is the yahoo description of this group:
Discussion of Color Theory and curves based color
correction in Photoshop ala Dan Margulis.
It seems that for some people the graphs and other
analytical tools
mean as much, or possibly *more* than the pictures they
are associated
with.
Absolutely. Key word: *more*. We're at a
picnic, with the opportunity to discuss color correction with Dan, and
everything's become covered in ants.
Anyone bring a can of Raid?
*** Discussion of Color Theory and curves based color
correction in Photoshop ala Dan Margulis. ***
Mike Russell
www.curvemeister.com/forum/
___________________________________________________________________________
Subj: RE: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Wednesday, September 27, 2006 6:18:12 PM
From: Randall.N.Hoffner
As I understand it, artificial lenses are made of
plastic, and this should enable you to see into the ultraviolet region
(unless they put some kind of filter in the material). I heard this
as far back as high school.
Randy Hoffner
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(Why the science?)
Date: Wednesday, September 27, 2006 6:18:29 PM
From: Richard Wagner
RJay,
You ask a very good, very important question, so I've
taken the time to give you a thoughtful answer.
How many images do you see/work on a year? If you
see 20 images/day * 5 days/week * 50 weeks/yr, you're looking at 5,000
images/year. If you see 200 images/day, that's 50,000 images/yr. If
the problem occurred in 1% (1/100), that would be 500 images with a
problem. Many photographers work with a lot more images, often using
automation. It is potentially a big deal. Then again, with
regards to the ongoing discussion, no one has shown that there is a problem
yet! There is a question, and the goal is to answer it so that we
know whether or not a problem exists, and if it does, when it is likely to
occur and whether there's a way to avoid it. So don't throw away your
valuable Lab tools just yet...
Picking a few "representative" images is not
a good test as to whether or not a problem exists. With this
approach, the odds are great that a problem that occurs with low frequency
will not be recognized. If 1 out of 10,000 new Ford F-250's leaked
oil and burned up the engine, would you likely find the problem if you
randomly picked 10 trucks on a Ford lot to examine? After doing so,
could you state with authority that the problem did NOT exist? A more
systematic approach is needed, and this is where the science comes in.
All of us would like our tools to work as we expect, all the time,
without the tools creating more problems for us, especially unexpectedly.
Let's say you enjoy working with wood, and you're going
to build a guest house. You mostly use a rip saw for framing the house and
building the roof, and it works great - over 1000 cuts without a problem.
Hey, the guys on the local construction project recommended the saw, and
they said it always cuts great. Now you go to make a fine hardwood table.
Same saw? You give it a try and it ruins your nice, expensive wood!
If you know that cross-cutting a nice hardwood board with a rip saw
tears up the edges of the board, when you go to make a table and you want a
nice fine cut, you'll switch saws, even though you'll continue to use the
rip saw for everything else (hey, it cuts fast!). It doesn't matter
if the other 1000 cuts that you made to make the roof were good enough, or
if the saw works great for the construction guys all the time. If you know
ahead of time that using a rip saw on a fine hardwood board is going to
ruin the cut, you'll change tools before the problem occurs, rather than
ruin the wood and waste your time and money. On the other hand, just
because the rip-saw may give bad results in a small percentage of your
work, it doesn't mean the tool is no good and it should be thrown out!
You're smart, and you can predict when the problem will occur, so that you
can use that excellent, fast-cutting rip saw most of the time, and the
fine-toothed, slow-cutting crosscut saw only when needed. The same is
true with imaging tools. A great tool may work well most of the time,
but not all of the time, and it's nice to be able to predict when it will
work well, or fail, even if it's a small percentage of the time that it
fails (or just creates a jagged edge).
I'm just a lowly prepress guy, but Dan's point that all
the theory and
mathematical proof in the world mean nothing if you
can't produce an
image that backs up the theory's predictions makes a
lot of sense to
me. I suppose its my lack of scientific training that
makes me think
that way.
Maybe. And it's my training as a
scientist/programmer that makes me question these things. Who
thought up all this digital stuff, and made it work? If not for all the
"math nuts" like Lindbloom who worked out all of this in the
first place, there would be no Photoshop, with its color spaces and
transformations and levels and filters and channels and histograms,
or any other digital imaging. You'd still be working with a sep house and
stripping film, which is where things stood when I worked as an offset
pressman. Back then, a densitomiter in the shop was a big deal.
Color theory was a necessary foundation for all of this "digital
workflow" to come to fruition, and it is still necessary for the field
to progress. (Remember Photoshop 4?) Be glad that there are those
with an interest in checking, fixing and improving your tools! You
have to understand the math to write the software that does all of the cool
things that we take for granted. But I see your point - perhaps this list
should have been called, "Practical Photoshop" or something other
than "Color Theory." So with that in mind, I'll bow out and
work on the problem off-list, then present my findings to the list in one
(hopefully coherent) post, and hopefully in a way that everyone can follow
(or hit delete). It isn't rocket science - it just takes different
tools to analyze the question, and the patience and time to carefully work
through it. Without studying the questions, I certainly don't know the
answers, and in my book, guesses don't carry a lot of weight.
Best,
--Rich
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: AdobeLightroom Beta)
Date: Wednesday, September 27, 2006 6:21:15 PM
From: john_denniston
Since this sRGB Lab debate is about theory rather than
what the eye can see I'm going add a spanner to the works with a quote from
the InterPARES Project that is relevant. This project, started about 6
years ago, is concerned with the authentication of original documents. The
relevant quote is as follows:
As the Preservation Task Force observed early on in its
deliberations, "strictly speaking, it is not possible to preserve an
electronic record. It is only possible to preserve the ability to reproduce
an electronic record. It is always necessary to retrieve from storage the
binary digits that make up the record and process them through some
software for delivery or presentation."
At the lecture I attended the importance of this
statement was explained to mean that every time a digital file is opened
and closed, even though no modifications are made to the file, the file
changes. Moving a file from one computer to another changes the file.
Moving it from one folder to another on the same computer, changes the
file. This is true be it picture, text file, or data base. The words of the
text file don't change, the data in the data base doesn't change but the
bits that describe it do, and these changes are easily measurable, and it
is this fact that makes it difficult to authenticate any digital document
as the original.
Which leads me to Rich Wagner's statement:
The purpose of ANY operation on an image is to change
the file - that part is obvious. Ideally, you SHOULD be able to make a
"round-trip" to another color space and end up where you started.
That is the whole principle of ICC Color Management.
What has been learned from the InterPARES Project is
that this is impossible. If just opening and closing a file will
change it in ways that are measurable then moving the file from one colour
space to another and back will definitely change it in ways that are
measurable. Does this mean the picture is different in any way meaningful?
I'd say no, just as the words in the text file can remain the same
but the file measured as changed, so the picture can remain the same and be
measured as changed.
Regards,
John Denniston
___________________________________________________________________________
Subj: [colortheory] Re: sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Wednesday, September 27, 2006 11:03:31 PM
From: Andrew Rodney
All the recent discussions easily fall into the
category of Color Theory!
Easy enough to remove that from the description if the
host so desires.
Andrew Rodney
http://www.digitaldog.net/
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: AdobeLightroom Beta)
Date: Wednesday, September 27, 2006 11:04:01 PM
From: Richard Wagner
John,
What you have stated is taken out of context, and used
in a manner that I highly doubt the original authors would agree with.
Digital authentication of files is unrelated to whether or not the
image data in Photoshop files changes with everyday use, or in
unpredictable ways, without someone using subterfuge to try to deceive us.
If you randomly change bytes, even by one bit, in an executable file,
the odds are good it will forever crash or misperform. If copying
files from one disk to another always resulted in changes to the file data,
our computing world would not exist as it does. Your extrapolation
has no basis in reality. As a test, copy a file, then open and close
it 10,000 times. Then compare the file to the original using a file
utility. There will be no difference. If there is, you need to
have someone check out your computer.
--Rich Wagner
___________________________________________________________________________
Subj: [colortheory] sRGB to LAB conversions (was Re:
Adobe Lightroom Beta)
Date: Wednesday, September 27, 2006 11:08:39 PM
From: Stephen Marsh
Lee Clawson wrote:
Up until this recent discussion we didn't know to even
pay attention
to the image after the move to LAB and back.
Dear Lee, the issue has been made known on this list
more than once in recent years (both theory and applied theory). I expected
this debate at the launch of Dan's book, not some time after!
Regards,
Stephen Marsh.
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Thursday, September 28, 2006 4:24:07 PM
From: Marco Ugolini
In a message dated Sep 27, 2006 2:08 PM, Mike Russell
wrote:
Absolutely. Key word: *more*. We're at a
picnic, with the opportunity to
discuss color correction with Dan, and everything's
become covered in ants.
Anyone bring a can of Raid?
Mike,
That is a truly obnoxious remark.
By the way, my heartfelt apologies if those of us who
attempt to probe into issues of COLOR THEORY (ahem!) are such an annoyance
to the likes of you.
---
Marco Ugolini
Mill Valley, CA
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Thursday, September 28, 2006 4:24:34 PM
From: Bob Frost
Randall,
Yes, the first intraocular lenses did allow blue light
and UV through, but for some time, the lenses have been modified to exclude
UV, but allow blue light through, as our natural lenses did when we were
young.
Now, however, there is new research that suggests that
the age-related yellowing of our lenses protects our retinas from macular
degeneration! So the latest lenses are now yellow (equal to the yellowness
of a 50yr old lens) to protect us from the blue light that we used to be
able to see when we were young.
So even at the relatively young age of 50, you are
looking at the world through a yellow filter.
So don't spend too long agonising over the correction
of a color containing blue. Those who look at your image will probably all
see different shades, according to their age!
Bob Frost.
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Thursday, September 28, 2006 4:27:03 PM
From: Bob Frost
Andrew,
All the recent discussions easily fall into the
category of Color Theory!
Easy enough to remove that from the description if the
host so desires.
I last suggested that almost exactly a year ago:-
"It is the "Colortheory" group, is it
not (although ColorPractice might have been a better choice of title for
Dan to use)." Sept 2005
Dan seems happy with the current title, as he
replied:-
You have brought up the point of the name of the list
previously, and Stephen
Marsh replied that it is called the *APPLIED* Color
Theory list. Over the
last two or three years, the conventional wisdom has
changed rather greatly as "experts" have discovered that how
color theory is applied and how it might work in a vacuum are very
different things.
Bob Frost.
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Thursday, September 28, 2006 4:46:07 PM
From: RJay Hansen
On 9/27/06, Mike Russell wrote:
Absolutely. Key word: *more*. We're at a picnic, with
the opportunity to
discuss color correction with Dan, and everything's
become covered in ants.
Agreed. Unfortunately, at times the signal to noise
ratio in this group is very low. Too many trying to pursue personal agendas
and/or vendettas.
RJay Hansen
___________________________________________________________________________
Subj: RE: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Thursday, September 28, 2006 6:16:46 PM
From: David Cardinal
Sorry, on thinking further I'm having two problems with
all
this: (1) the item at issue is loss of colours due to
conversion from one colour space to another. It isn't
clear
to me how a demonstration of what happens when altering
curves is relevant to what happens when you convert a
file
from colour space A to colour space B.
My point is that if even a fairly simple everyday
operation like applying a curve has the same effect of eliminating large
numbers of potential unique colors, then that fact alone doesn't seem to be
enough to me to condemn a colorspace conversion that happens to have the
same side effect.
--David Cardinal
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Thursday, September 28, 2006 6:22:06 PM
From: Terry Wyse
Marco, you keep pressing on despite the silly remarks.
I for one am very interested in your findings. It's kind of a shame that
the less-than-civil tone adopted by the moderator seems to be infecting the
hoi polloi. It only causes to increase the noise level (and waste
bandwidth) and further distracts us from getting to the truth, wherever
that may take us.
I would argue that if in fact what you're positing is
true, this takes it from color THEORY to REALITY.
;-)
Regards,
Terry Wyse
_____________________________
WyseConsul
Color Management Consulting
G7 Certified Expert
704.843.0858
http://www.wyseconsul.com
http://www.colormanagementgroup.com
___________________________________________________________________________
Subj: [colortheory] Tips on LAB Workflow
Date: Thursday, September 28, 2006 6:26:57 PM
From: David Cardinal
With all the concern over possible issues with
converting to LAB & back, I thought I'd pass along the workflow I use
that in addition to being quite quick also sidesteps most of these issues:
1) I use list member Mike Russell's awesome plug-in
Curvemeister to do LAB curves. Since 99% of what I want to do in LAB can be
done in Curves, this tool allows me to do all the work from RGB and allow
Mike's code to do the heavy lifting. This way I can keep all my layers,
etc.
2) I always do the Curve on a separate layer
(unfortunately the LAB curve can't just be an adjustment layer, which would
be even cooler), which has a few benefits:
-> I can vary the opacity to control the effect.
This probably has an interesting effect of reducing any impact of the lab
conversion since I'm partially using the RGB colors.
-> I can of course turn off the layer and at any
time have all my original RGB pixels in case I want them or want to have
them pre-LAB effect.
-> I can mask the LAB effect as needed as well, if I
want to enhance only part of an image.
I hope this is helpful. I believe Curvemeister is
win-only, although there might be something similar for the Mac. I've
written up some of this in more detail:
http://www.nikondigital.org/dps/dps-v-4-05.htm
--David Cardinal
Cardinal Photo / Pro Shooters LLC
http://www.cardinalphoto.com <http:
//www.cardinalphoto.com/>
http://www.nikondigital.org <http:
//www.nikondigital.org/>
http://www.proshooters.com <http:
//www.proshooters.com/>
___________________________________________________________________________
Subj: RE: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Thursday, September 28, 2006 6:30:11 PM
From: David Cardinal
I agree with you in principle, but not in detail. I'm
not
sure where you get 215*215*215. I get
(256-20)**3, or
13,144,256, which can be confirmed by counting the
number of
unique levels in the resulting file with the Levels
app.
This results in a loss of 3,632,960 unique colors from
the
original 16,777,216 unique colors in this image. As you
describe, the loss of levels is quite easy to see on an
RGB
histogram, even in PS CS2 (try it!); there are 20 sharp
"spike canyons" on the left where the levels
are lost, and 20
"spike peaks" on the right where the
resulting levels of changed pixels
overlap pre- existing levels, raising the frequency
count
from 1 to 2. (I used a +20 change.)
Yes, that's probably more correct. I was using 40 since
20 of the values on the left are now mapped onto adjacent values so the
information they contained is lost (a loss of real colors) but other values
are mapped onto their original values so it doesn't cost us theoretical
colors. The point remains the same though.
As to your point about real vs. synthetic., Amen. I
personally don't think the unique color statistic is very helpful in this
case since all the numbers for the LAB conversion are over 2M, which sure
seems like a lot:-) I'm much more interested in a model which provides some
measurement of the amount of visible data loss for a transform.
Presumably by definition visible data loss would be
visible, so to Dan's point we should be able to see it if it is important.
I haven't seen it in doing simple conversions, but I do wonder if I convert
a file and then tinker with it (sharpening, etc.) whether the cumulative
effect starts to do some damage. --David Cardinal
___________________________________________________________________________
Subj: [colortheory] sRGB to LAB conversions (was Re:
Adobe Lightroom Beta)
Date: Friday, September 29, 2006 7:21:35 AM
From: Olivier Desmaison
Presumably by definition visible data loss would be
visible, so to
Dan's point we should be able to see it if it is
important.
That's the point.
For what I understand, data loss would not necessarily
*be visible* because the evaluation is not to be done on the value codes
but on the gamut volume possibly based on DeltaE cubic, so you need an
additional fonction to be performed on the value codes to get the transform
into colors and then the computation of variation. Please see :
http:
//www.media.hut.fi/~as75192/tenttima/Ext_gamut.pdf#search=%22gamut%
20volume%20calculation%22
and
http:
//ci.uofl.edu/tom/papers/Cholewo99cic-titled.pdf#search=%22gamut%
20volume%20calculation%22
Olivier Desmaison
___________________________________________________________________________
Subj: [colortheory] Endless arguments (was at one time
something about color)
Date: Friday, September 29, 2006 7:26:52 AM
From: J Walton
On 9/27/06, Marco Ugolini wrote:
Mike,
That is a truly obnoxious remark.
By the way, my heartfelt apologies if those of us who
attempt to probe into issues of COLOR THEORY (ahem!) are such an annoyance
to the likes of you.
I don't know if this will make it past a moderator or
not, but it should. I happen to disagree. I think Mike's point is perfectly
valid, and here's why:
We all know what this list is about. It's not just
about community, or just a group of professionals who like to talk about
color theory. It's a list provided by Dan for people who have either bought
one of his books or paid for one of his classes. It's not that hard to
understand - he reminds us of it every month.
For some reason there's a vocal minority who seem to
spend time here just so they can get into arguments with Dan. And for
whatever reason Dan is moved to push their buttons is such a way as to get
an amazingly predictable response. It goes like this:
1. Someone brings up an innocent topic (Lightroom,
Epsons, the price of tea in China). Somehow the conversation swings around
to 16-bit or ProPhoto and the word clipping or quantization appears.
2. Dan chimes in and wants to see actual images to back
those assertions.
3. Someone replies back and accuses Dan of being an
unscientific hack, usually including some mathematical reference.
4. Dan replies back and uses other mathematical terms
and accuses the others of never having studied the subject. Usually at this
point someone is accused of taking a payoff a decade ago.
5. The same someone everytime replies back and wants
Dan to provide check stubs to prove that someone was paid to change their
views. That's when Dan checks out of the conversation.
6. The thread continues ad nauseum until a moderator
requests that it be moved to the ColorSync list or ended completely.
7. A few weeks later, step one happens all over again.
DO WE REALLY HAVE TO GO THROUGH THIS OVER AND OVER
AGAIN?
The target audience for Dan is people who find his take
on color correction refreshing, and who want to hear more. I happen to be a
part of that audience, and I've been reading his columns for well over a
decade.
Make no mistake, I don't agree with everything Dan
preaches. And I know there's things I can do in Photoshop that he can't. I
read this list because I want to get better at what I do, and because Dan's
take on color correction has helped me get this far.
As much as Andrew and Chris and plenty of others (I
can't name you all) REALLY know their stuff and are successful writers in
their own right, nobody's joined their email list. And if there was such a
thing and someone signed up merely to poke holes in what they were trying
to do they would rightly be annoyed. And so would I.
And that's why I think Mike's post is entirely fair.
"We're at a picnic (email list), with the opportunity to discuss color
correction with Dan, and everything's become covered in ants (endless posts
about imaging trivia)."
A person who is *more* concerned about levels or big
numbers or histograms than they are about how the image looks is not
necessarily wrong. They're just in the wrong place.
-----
J Walton
___________________________________________________________________________
Subj: Re: [colortheory] sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Friday, September 29, 2006 12:11:28 PM
From: Andrew Rodney
Visible on what device (and when)? Lots of stuff is
invisible on a display. You can usually find banding on many LCD displays
due to the 8-bit graphic system and such banding is NOT in the file. Plus
you1re limited to seeing anything outside sRGB. Halftone dot? You can
output stuff that will look fantastic only to see issues with a good Epson
at 2880 output. How about output to a true contone device like a Lightjet?
Anyone remember film recorders? I know folks still using them today.
Andrew Rodney
___________________________________________________________________________
Subj: Re: [colortheory] Endless arguments (was at one
time something about color)
Date: Friday, September 29, 2006 12:16:47 PM
From: Andrew Rodney
On 9/29/06 1:13 AM, "J Walton" wrote:
1. Someone brings up an innocent topic (Lightroom,
Epsons, the price
of tea in China). Somehow the conversation swings
around to 16-bit or
ProPhoto and the word clipping or quantization appears.
Based on your criteria of what this list is all about,
that topic should not have made it in the first place.
Andrew Rodney
___________________________________________________________________________
Subj: [colortheory] Re: Endless arguments (was at one
time something about color)
Date: Friday, September 29, 2006 12:18:04 PM
From: George Machen
This nonsense has been going on for years and I am sick
and tired of it. This is a moderated list. Perhaps the moderators would
consider ending these annoying distractions once and for all by enforcing
off-topic restrictions on trolls. One way to do it would be to permit one
such post, with a polite redirection to the ColorSync list.
(It will not hurt my feelings if this message doesn't
get posted - it is itself off-topic and mainly a heartfelt request to the
list moderators.)
George Machen
___________________________________________________________________________
Subj: Re: [colortheory] Endless arguments (was at one
time something about color)
Date: Friday, September 29, 2006 12:24:58 PM
From: Bob Frost
I hope my posts on the variation in color vision in
humans aren't regarded as trivial. This variation seems to be ignored by
most people involved in color correction. I dug up a graph yesterday from a
lens manufacturer showing the difference in color transmission of a
25yr-old human lens and a 54yr-old lens. Here are a few figures to show the
difference:
wavelength 450
500 550
600 650 700
25yr-old 35%
70% 75%
75% 75% 75%
54yr-old 25%
45% 50%
55% 57.5% 60%
Reduction 29%
36% 33% 26%
23% 20%
As you can clearly see the older lens has 20% reduced
red transmission, but this progressively increases across the spectrum to
36% in the blue. And this difference is found at a mere 54 yrs of age! It
is much greater at 60, 70, and 80.
Food for thought! And this is only one of the
differences in color vision amongst us.
Bob Frost.
___________________________________________________________________________
Subj: Re: [colortheory] Endless arguments (was at one
time something about color)
Date: Friday, September 29, 2006 12:28:23 PM
From: Ron Kelly
On 29-Sep-06, at 1:13 AM, J Walton wrote:
A person who is *more* concerned about levels or big
numbers or
histograms than they are about how the image looks is
not necessarily
wrong. They're just in the wrong place.
Well said.
Ron Kelly
___________________________________________________________________________
Subj: Re: [colortheory] Endless arguments (was at one
time something about color)
Date: Friday, September 29, 2006 3:18:43 PM
From: Lee Clawson
The problem with the can of Raid is how easy it is to
exterminate some good ideas with the annoying pest-like ones. Can you be
sure where things will lead and who among will have and share an insight
that makes things clearer and easier for all of us ???
Lee Clawson
2/\V/\7 Studio
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Friday, September 29, 2006 3:23:29 PM
From: Mark Segal
Marco, you're right - intellectually it would be a good
learning experience for me to replicate your test, but I'm up to my
eyeballs, so I need to make decisions about priorities and trust things I
think I should trust and do other things that only I can do.
I think the result you got on the 16 bit file is
an extremely important piece of information for this whole debate - puts a
number of things in a whole other perspective, doesn't it? (Which is why I
asked.) Indeed, the Lab conversion does turn out to be much more
"gentle" and the difference before and after the conversions in
both cases would be completely undetectable in a print.
Cheers,
Mark Segal
___________________________________________________________________________
Subj: Re: [colortheory] Endless arguments (was at one
time something about color)
Date: Friday, September 29, 2006 3:25:34 PM
From: Henry Davis
On Sep 29, 2006, at 3:13 AM, J Walton wrote:
And that's why I think Mike's post is entirely fair.
"We're at a
picnic (email list), with the opportunity to discuss
color correction
with Dan, and everything's become covered in ants
(endless posts about
imaging trivia)."
A person who is *more* concerned about levels or big
numbers or
histograms than they are about how the image looks is
not necessarily
wrong. They're just in the wrong place.
I agree - this kind of discussion is better off taking
place on the other list. When I responded earlier with a post that
would bring printers and inksets into the relevance of the discussion, I
hoped that it might swerve the discussion in a direction that we might
could benefit from. To his credit, Andrew posted the only response.
The on-goings within the workings of a color space profile, color engines
and rendering intents are interesting, but regardless of the hows and whys,
we are nonetheless left with an image. Discussions about these
aspects are fine as long as they're offered in an attempt to illustrate
their effects on real images. The image is where I spend my time.
When participants totally disregard the image as though the image is
OT, that's when I've had enough. I'm not a developer, but if I were,
I would spend time listening to this list for input, and not use it for
endless talk about integer loss and rounding errors. I'm happy to
follow along with the arcane discussions when and if they get around to
relevance to real pictures. So far, this one has all but condemned
the picture to irrelevance.
Henry Davis
___________________________________________________________________________
Subj: Re: [colortheory] Endless arguments (was at one
time something about color)
Date: Friday, September 29, 2006 3:26:28 PM
From: Henry Davis
On Sep 29, 2006, at 9:19 AM, Bob Frost wrote:
I hope my posts on the variation in color vision in
humans aren't regarded
as trivial. This variation seems to be ignored by most
people involved in
color correction. I dug up a graph yesterday from a
lens manufacturer
showing the difference in color transmission of a
25yr-old human lens and a
54yr-old lens. Here are a few figures to show the
difference:
I don't consider your post to be trivial, and as food
for thought, it seems to me that there must be some adaptation at work in
our mind's eye. The reduction of color transmission that these
numbers suggest is very big, if only the numbers are considered.
Maybe it takes 54 years of constant adaptation to correct for the
aging process of the lens. Uh oh, now we've swerved dangerously close to
the controversial topic of evolution. Oh, no, now I've done it - it
isn't considered controversial to some, while for others it's open for
debate.
Henry Davis
___________________________________________________________________________
Subj: Re: [colortheory] Re: sRGB to LAB conversions
(was Re: Adobe Lightroom Beta)
Date: Friday, September 29, 2006 3:26:33 PM
From: Lee Clawson
on 9/28/06 11:54 AM, David Cardinal wrote:
but I do wonder if I convert a file and then tinker
with it (sharpening, etc.)
whether the cumulative effect starts to do some damage.
To me that's the point. We need to watch for the
effect(s).
With Dan's articles & book highlighting LAB edits I
was surprised to see that simply moving from sRGB to LAB and back without
doing any edits gave a loss.
Lee Clawson
2/\V/\7 Studio
P.S.-- (sorry Stephen I really did miss the old posts
that pointed this out)
___________________________________________________________________________
Subj: [colortheory] Why I'm on this list
Date: Friday, September 29, 2006 3:27:53 PM
From: john_denniston
There are a number of people who contribute frequently
to this list who seem to be here for the sole purpose of proving that Dan
Margulis is wrong.
I joined the list in 1996, a year after the paper I
worked at bought 22 NC2000 cameras from AP, closed its film lab and went
100% digital. To pay for the cameras it fired all the engravers who knew
anything about color reproduction, assuming that as photo editor I would
learn the necessary Photoshop tricks in a few weeks and be able to take
care of that part of the business.
Of course we had a lot of problems so we brought in
consultants. Most of them concentrated on providing profiles to our four
presses. None of them succeeded in getting the colour better or anywhere
near an acceptable quality. Someone told me of this listserv and Dan's
books which I bought and read. As I applied his theories to problems, they
were solved. I once brought up Dan's name with one of the consultants and
he told me that Dan was a "dangerous man". After spending two
weeks at the paper this consultant left having sold us some very expensive
calibration equipment that didn't improve our colour reproduction one iota.
By the time I retired 3 years ago the colour
reproduction at the paper was very good, possibly the best in the chain, an
accomplishment difficult to achieve considering the company's insistence
that the people doing pre-press were to be few and basically untrained in
colour correction. A lot of people made it happen and Dan, indirectly
through this list and his books, was one of them.
Opinions on this list are free. I rate their value as
to how well they work, do they help solve my problem. Lots of people have
provided advice to me over the years but some of it has been faulty, it
hasn't provided a solution to my problem or worse, headed me in the wrong
direction, and I learned not to rely on advice from these people. I learned
to recognize good advice and ignore the bad. It's something a couple of
people on this list should do more of. Dan's opinion of LAB and 16 bit is
unlikely to change and it shouldn't bother us. He offers his theories to
help solve problems. If they work, use them. If they don't, ignore them.
I'm still here because Dan's advice and that of others
has always helped. He has been accused of being gruff and insensitive in
offering his opinions. Anyone who walks into a pressroom to tell the
foreman he has to do things differently had better be. I know, I've been
there. If this is a problem for you, you're going to have trouble surviving
in this business.
Regards,
John Denniston
www.dennistonphoto.com
www.dirtbikephoto.com
___________________________________________________________________________
Subj: [colortheory] data loss in Lab conversions?
Date: Friday, September 29, 2006 3:35:08 PM
From: Ronald Greenburg
I'm writing this as a new post instead of a response to
anyone in particular because I don't want to offend anyone in particular.
The arguments going back and forth have ranged from the informative to the
ridiculous, maybe also to the sublime. One problem seems to be that two
separate arguments are going on; lots of you people are talking past each
other rather than at each other; you are trying to prove two completely
separate things, but most frustrating of all, you are trying to convince
each other whether something 'matters', and since both camps have an
entirely different basis for judging whether something 'matters', you are
all obviously condemned to fail in your attempts. Thank God we are just
having fun talking about how to make our images better (applied color
theory), and not married to each other!
Visible data loss is visible. It's not necessarily
'data loss', even if one can show a decrease in the number of 'discrete
colors'. That presumes the changes to the bits represents 'data'. Maybe it
represents noise. So if we run a filter or program to reduce noise in an
image, have we 'harmed' the file? Of course not, even though we probably
reduced the number of colors.
Nothing can 'permanently harm a file' and at the same
time be invisible to a group of trained observers, or even to yourself. For
two reasons. First of all, I presume no one is stupid enough to throw away
their original raw file or other original capture. So if you don't like how
you have changed a file, you just go back and start again. We have all done
things that have 'damaged' a file. I know I have, it was when I applied a
curve or did some other move that made my picture look worse instead of
better. So you go back and undo or start over.
Second of all, what is the point of going from one
colorspace to another, except to apply a curve or do a maneuver which you
can't do in the first colorspace, with the intent of improving your image
thereby? What moron wants to go from one colorspace to another and back,
purely for the fun of making the trip?
Which leads me to ask, why is all this arguing going on
in the first place? Either the arguments about data loss, quantization
error and image degradation are to try and convince you not to use Lab, or
else they are not. If they are to try and convince you not to use Lab, then
just read Dan's book. He will show you wonderful things you can do to
improve images, and things which can be done in no other way than by using
Lab. If these arguments are not about that, then what the hell are they
about? I am sure there are some noble academics amongst you who are
interested in the numerical arguments for some purely theoretical issue
besides the question of actually using Lab or not. For you I say, carry on,
wonderful people.
But for those who are less well-intentioned (and I
can't tell whether any of you might fit into this category), I can simply
say, your arguments are neither right nor wrong, but merely irrelevant. As
Dan might say, so what? What is the point? I would rather listen to people
talk about new ways to use Lab, maybe when to not use Lab. However good Dan
might be, I cannot imagine that he has emptied the well. I would like to
hear about new ways to use Lab. Hasn't anyone thought of ways to use Lab
creatively besides Dan? This would be worth talking about.
I was taught that the scientific method consisted of
testing theory by empirical facts. One camp of you, I will call it 'pure
color theory' is looking at numbers. Another group, I will call 'applied
color theory' is looking at photo images. My conclusion is that pure theory
does not lead in any way to applied theory; they may use similar jargon but
have nothing to do with each other.
Thank you for this brief intermission ... Let the games
continue !!!
Ronald Greenberg
___________________________________________________________________________
Subj: [colortheory] sRGB to LAB conversions (was Re:
Adobe Lightroom Beta)
Date: Friday, September 29, 2006 3:40:24 PM
From: Olivier Desmaison
The way I see it : The Lab color correction goes
together with other issues such as color space and bit depth with a final
output mostly on press.
Put all things together : color space is sRGB, given
sRGB the round trip only generate minor deviations, and 8b suffice for the
file. Last, given most of the corrections are based on human decison (
*choices* or "feelings" the colors should be more saturated,
greens should pop up...), the Lab corrections that you actually *see* on
screen (display profile and sRGB come close) and that will utimately be
your sRGB file make somehow irrelevant whether there are variations or not.
The last argument will always be "you don't see it" plus anyway
you will output it more or less the way you see it.
I am interesting in assessing those variations, but
given the whole workflow (consistent and coherent as it is again given the
context), I also understand there are considered meaningless.
The choice of the working space tells a lot on what
device your workflow is based on : Prophoto->DSLR, WideGamut->scan,
aRGB->desktop printer, sRGB->display/press. Apple even mentionned
that their choice for keeping 1.8 gamma was directly related to the
reproduction of the press low contrast capability.
Yet, for discussion sake, even if it's only
theoretical, I favor the hosting by this list of the topic : it would be a
pity Lab has another trip to go to another list to be discussed.
Olivier Desmaison
___________________________________________________________________________
Subj: Re: [colortheory] Why I'm on this list
Date: Friday, September 29, 2006 5:15:33 PM
From: Jim Donovan
Opinions on this list are free. I rate their value as
to how well they work, do
they help solve my problem. Lots of people have
provided advice to me over the
years but some of it has been faulty, it hasn't
provided a solution to my
problem or worse, headed me in the wrong direction, and
I learned not to rely
on advice from these people. I learned to recognize
good advice and ignore the
bad. It's something a couple of people on this list
should do more of. Dan's
opinion of LAB and 16 bit is unlikely to change and it
shouldn't bother us. He
offers his theories to help solve problems. If they
work, use them. If they
don't, ignore them.
I'm still here because Dan's advice and that of others
has always helped. He
has been accused of being gruff and insensitive in
offering his opinions.
Anyone who walks into a pressroom to tell the foreman
he has to do things
differently had better be. I know, I've been there. If
this is a problem for
you, you're going to have trouble surviving in this
business.
Very ,Very well said. After being a journeyman
stripper, running a drum scanner for eons and now a scitex..Kodak Eversmart
pro for as long as they have been out there is no question what has to
happen for sucessful offset reprodution. The fact is you can buy and apply
all the "stuff" you want or you can learn to make good seps. The
Alpha and Omega are good seps,game over. Some people will deny this reality
all day and insist that buying something can replace knowledge of offset
printing.Learn the process like we did and you will wil be fine. Isn't it
amazing nobody ever adresses the galring fact there would be a big void in
color reproduction when color houses went in house and designers were
expected to pick up the slack. Insted of learning people have been trying
to find a magic bullet that dose not exsist without a learning...not
tagging.... process. Jim Donovan
___________________________________________________________________________
Subj: RE: [colortheory] Why I'm on this list
Date: Friday, September 29, 2006 5:17:42 PM
From: Sterling Ledet
I can second the opinion that Dan is a dangerous man.
Particularly to those who are in the business of selling people hardware
and software as magic bullets that can prevent people from having to learn
the fundamental concepts of color correcting by the numbers.
He's also dangerous to those who think they can out
drink or out eat him as I've unfortunately learned the hard way. I
understand there are quite a few casino operators who view him as a
"dangerous man" as well.
So at least that's one bit of truth the consultant was
probably dead on with. It's just that Dan's isn't likely to be dangerous to
"you"!
- Sterling Ledet
___________________________________________________________________________
Subj: Re: [colortheory] Re: Endless arguments (was at
one time something about color)
Date: Friday, September 29, 2006 8:20:04 PM
From: Alan Shils
This nonsense has been going on for years and I
am sick and tired of
it. This is a moderated list. Perhaps the
moderators would consider
ending these annoying distractions once and for
all by enforcing
off-topic restrictions on trolls. One way to do
it would be to permit
one such post, with a polite redirection to the
ColorSync list.
George, sometimes some people want to see how dead a
dead horse can be beaten deader.
.
.....Alan Shils
___________________________________________________________________________
Subj: RE: [colortheory] sRGB to LAB conversions (was
Re: Adobe Lightroom Beta)
Date: Friday, September 29, 2006 8:20:18 PM
From: David Cardinal
Cool papers. From flipping through the first one it
seems to conclude that LAB 8-bit doesn't have any more quantization error
than sRGB, which is good to know.
To put this conversation in perspective, here is what I
think is ultimately important to me and I suspect to many others:
* I "process" & print a lot of my photos,
many at large sizes. This is a big piece of how I make my living. Printers
are either large format Epsons or Lightjets.
* Other photos go off to editors and go through
whatever backend process they have for their publication. Frankly I don't
worry too much about those because aside from supplying a good quality
original in my best estimate of the "right" colorspace for the
client, I don't have much control.
* For the photos I process & print, I'm seeing some
really great results by using some of the LAB techniques I first learned
from Dan--and some of my own that I've developed since.
* The quality of my prints has improved as a result,
but of course I'm never 100% happy. So in cases where I want to improve the
prints further, the natural question is whether the LAB techniques I'm
using can in turn be improved in some way, or if they cause any issues
which could be addressed by doing thing differently.
So, my interest is strictly in deciding what techniques
are the best for what I do and where I should be looking to further improve
on them.
So far I haven't seen anything to make me stop doing
LAB corrections when they (to my eye) improve my images, but I am on the
lookout now to see if I find any concerning side-effects. In particular it
is super-easy to over-saturate colors by doing moves in LAB which can
destroy detail big time. So that's something I pay a lot of attention to
when I use LAB as a tool.
--David Cardinal
___________________________________________________________________________
Subj: [colortheory] Theory vs. Practice
Date: Friday, September 29, 2006 11:06:55 PM
From: Howard Smith
With references to the recent posts about endless
arguments, while half of the title of this forum is "Theory", the
only real value to me over the years has been the discussions of practical
problems. John Ruttenberg's recent posts set me off on a search that
resulted in my finding Raw Shooter and Adobe Lightroom Beta 3 (now Beta 4),
download Adobe Raw (the old version didn't work on Olympus raw files), and
stimulated me to look more deeply into channel blending, etc. After
reading Jerry Fusselman's post about the hard shadow in a portrait, I spent
a day and a half learning new ways to lighten shadows. One of the
photographers told me years ago how to kill serious casts and now there
isn't a cast alive that I can't bag with a single shot. Others
wisely sought to convince me that going digital was the only reasonable
option when I was researching the possibilities of going into medium format
film photography. They were right, of course, and the recommendations
not only saved me from a considerable waste of money but pointed me
to a much better way of doing things. But more recently there have
been so many detailed forays into the mathematics of everything Photoshop
that it's becoming bothersome trying to keep my Inbox emptied fast enough.
I've even learned that at my age I can no longer distinguish colors,
even though the colors that I would swear I do see look remarkably like
those of long-ago memories.. It's not that all these things that are
way over my head aren't fascinating, but then so is reading about the
latest findings in particle acceleration. Neither one does very much
to pique any serious interest or to make me better at color correction.
I do admire the intellectual prowess of those who write the endless
analyses of why one method is better than the others, but all my sincere
admiration for the genius of many contributors still doesn't do me a bit of
good. Of course this is a selfish way to look at it, and I would
never want to discourage anyone from expressing their heartfelt opinions.
It's just that while I don't know about the others, it really doesn't
help make my work any easier.
Photoshop can be both fun and profitable if we work at
keeping it that way. If necessary perhaps Dan will be willing to break the
list up into Theory and into Practice, with neither one excluding any posts
or discussions that cover both topics. Right now I would really
welcome some practical posts from Mike Russell, John Ruttenberg, Stephen
Marsh and the many others who present down-to-earth problems and solutions
that have practical as well as educational value. Am I the only one
who think it is these kinds of things that make this forum fun? We
can find tips and tricks and better ways to create artistic fonts with
brushed metal textures in Photoshop User magazine, but we can't find
anything like the solid information that often passes through here.
Or that used to pass this way.
Agitated replies never offend me, so any of you who
strongly disagree are encouraged to fire away.
Howard Smith
___________________________________________________________________________
Subj: [colortheory] Re: data loss in Lab conversions?
Date: Saturday, September 30, 2006 10:56:26 PM
From: George Harding
Amen to that.
Many of the concerns raised about the changes
that occur with color space conversions, along with the arguments for use
of wide gamut color spaces and 16-bit files, are that even if you can't see
the difference with current displays and printers you may be able to with
future technology. This ignores the improvements in software that
will also occur down the road - just look at some of the raves about how
much better the raw conversion capabilities of Lightroom are as compared to
ACR3. And it won't be long before Adobe will be looking for us to pay
to upgrade to a new version of Photoshop with some new "must
have" features.
Probably more important, it ignores the
improvements in our ability to use these software tools. I've learned
a lot from the experts who share their knowledge with this group, and even
more as a beta reader for Dan's update of Professional Photoshop. I
can make corrections now I couldn't a year ago, and I expect in a year I'll
be able to do things I can't now. When the new display technologies
arrive, I'll use the improved software tools and my improved skills to get
the most out of them I can.
With regard to the extended p*ssing matches,
I wouldn't advocate a complete ban. In small doses it can be
entertaining and even educational. If it starts to irritate me I just
scroll on by.
George Harding
Marlborough, MA
___________________________________________________________________________
Subj: [colortheory] sRGB to LAB conversions (was Re:
Adobe Lightroom Beta)
Date: Sunday, October 1, 2006 10:14:20 AM
From: Olivier Desmaison
From flipping through the first one it seems to
conclude that
LAB 8-bit doesn't have any more quantization error than
sRGB, which
is good to know.
Yes given a constrained sRGB gamut for a real world
target and keeping in mind some real world colors will not be encoded.
Which naturally leads you to considered the output device gamut(s) as your
next gamut target to test.
Olivier Desmaison
___________________________________________________________________________
Subj: Re: [colortheory] Endless arguments - now
aging eyes
Date: Monday, October 2, 2006 9:12:33 AM
From: John Castronovo
Were they measuring people's perception of color or the
lenses themselves? I would think that the mind would adapt and compensate
for the physical changes in the lens tissue to a great degree. After all,
vision takes place in the mind and not the eye itself. We all have
automatic white point balance compensation.
john castronovo
___________________________________________________________________________
Re: Why I'm on this list
Posted by: "John R"
Thu Oct 5, 2006 11:09 am (PST)
I too am sick of the arrogant carping on Dan on this
list. I began reading Dan's articles at around Photoshop 4. the list a
little later. He's always asked for image files to prove him wrong, and all
the critics come up with is endless theory statements. I hope some day that
they please go away and set up their own list. I'm sure their supporters
will follow if there are any. However, they never seem to get it through
their heads to put up or shut up. They can't shut up. They have crap to
sell and by golly they're gonna try and unload it here. I know this is OT,
but I'm really sick of seeing their input year after year and the same
theoretical crap to unload.
John Robinson
___________________________________________________________________________
Posted by: "Richard Wagner"
Sun Oct 8, 2006 7:32 am (PST)
Quantization and Rounding Errors Associated with 8-Bit
Lab Colorspace
Conversions (Change of topic - was [colortheory] A case
of clear 16 bit superiority)
I am not surprised by Ric's observation that converting
his 8-bit image to 16-bit prior to the conversion to Lab resulted in much
smoother tones, with less noise.
The quantization and rounding errors associated with an
8-bit vs. 16- bit workflow involving conversions between sRGB (or other RGB
color spaces like ColorMatch) and Lab are easy to demonstrate. If an image
begins in sRGB and is converted to Lab, and then is converted back to sRGB,
there should be no change in the image. What you end up with should be
exactly what you start with. Any change in the resulting image is a result
of quantization error or rounding error.
By starting with an image containing all possible sRGB
colors in an 8- bit image, it can be demonstrated that converting the image
to Lab, then back to sRGB significantly alters the original colors in the
image - i.e., the process is lossy. In contrast, taking the same starting
image, converting it to 16-bit, then converting to Lab and back to sRGB,
then to 8-bit, results in an essentially identical image compared with the
original 8-bit image.
It is easy to show that starting with the 8-bit image
(256 levels per RGB channel), converting to Lab, then converting back to
sRGB results in errors of up to 24 levels in the reds, 9 levels in the
greens, and 10 levels in the blues. Taking the exact same starting image,
converting it to 16-bit, then converting to Lab and back results in an
error of, at most, one level per R, G, B channel.
The primary reason for the error in conversion of the
8-bit image to Lab and back is "quantization" error. 8-Bit Lab
does not have adequate "encoding" space for all of the source RGB
colors (i.e., the grid is too coarse), resulting in multiple source colors
getting lumped together as one Lab value. On conversion back to sRGB, all
of the original colors cannot be regenerated. With 16-bit Lab, the
"grid" is much finer. In 16-bit Lab space, the original 8-bit
sRGB colors will have lots of Lab encoding space between each color, the
encoding error is minimal, and regeneration of the original 8-bit colors is
possible.
The error associated with the 8-bit conversion is easy
to see visually using Photoshop. Start with an 8-bit image that contains
all 256x256x256 = 16+ million possible RGB colors (http://
www.brucelindbloom.com/RGB16Million.html). Open this in PS, assign the sRGB
profile, and save the file as "original." Convert this image to
Lab, then back to sRGB, using Convert to Profile (relative colorimatric,
+BPC). Save this as "Round-trip." Now use Photoshop to subtract
the original image from the round-trip image. If there have been no errors
generated by the color space conversions, the images will be identical, and
the image obtained by subtraction will be completely black (0,0,0). The
subtraction can be done in Photoshop CS2 by opening both images, then using
Image-->Apply Image...-- >Source=original file, Target=Round-trip
file, Blending=Difference. Save this image as Difference.tif.
The subtraction will have yielded a non-black pixel for
any changed pixel. If the error is minimal, the pixel will be close to
black (e.g., 0,0,1). The greater the error for a given pixel, the further
away its color will be from black (e.g., 15,5,3). To look at the Levels for
the Difference image, go to Image->Adjustments->Levels. Slide the
white point down to 25 or so. The errors in each of the channels can be
observed. Note that, when looking at the image and not the histogram, the
errors are distributed across all colors in the original image.
Repeat the above process by starting with the 8-bit
image, converting to 16-bit, and converting to Lab and back to sRGB, then
convert to 8- bit. The maximum error will be one level in each of the three
RGB channels.
Note that this has nothing to do with having additional
precision in a 16-bit image - e.g., 14 or 16 bits of precision vs 8-bits.
The errors shown are strictly from making a conversion to Lab and back in
8-bit mode.
--Rich Wagner
___________________________________________________________________________
Posted by: "Dan Margulis"
Sun Oct 8, 2006 3:06 pm (PST)
Re:Quantization and Rounding Errors Associated with
8-Bit Lab Colorspace Conversions
Rich Wagner writes,
The quantization and rounding errors associated with an
8-bit vs. 16-
bit workflow involving conversions between sRGB (or
other RGB color
spaces like ColorMatch) and Lab are easy to
demonstrate.
This is substantially the same statistically illiterate
argument that was posted two weeks ago. In response, I asked if you were
able to show a photograph that exhibited this damage. Many members of the
list chimed in, correctly demanding to see something more than histograms
and gradients to justify these conclusions.
The moderators have discussed the threads that took
place in my absence last week (prior to this particular post) and we all
agreed that there is no further room for "proofs" of this nature
when it is easy to prove or disprove the point with real photographs. This
particular issue has been done to death over the years, see, e.g.
www.ledet.com/margulis/ACT_postings/ColorCorrection/ACT-LAB-damage.htm
These conversions have been tested extensively. Nobody
can show any damage with real images even under unbelievably extreme
conditions, such as (with images specially chosen to be sensitive to
damage) converting and reconverting 100 times, or converting and
reconverting multiple times at different phases of the correction.
We have been hearing this "data loss" inanity
at least since 1994, when one Photoshop expert announced that neither CMYK
nor LAB should ever be used in color correction because the conversion
inflicted "catastrophic damage" on the image. His
"proof"? The same idiotic "levels loss" demonstration
that we still hear about today, which demonstrates only that colorspaces
have different gammas.
Your "proof" is similar. You take for granted
that some normal effect of conversion is wrong, and then because the
conversion behaves normally, you announce that you have proven that there's
damage. This is just what you did two weeks ago--you invented an irrelevant
statistic, "unique colors", and then announced as a given that it
indicated damage. This is more of the same: you announce as a given that
any variation in any channel is "damage" and proceed from there.
As a result, the rest of the post is worthless.
Statistical process control is tricky, and fools many
experts. You, on the other hand, even after Lee Clawson was kind enough to
link the definition, don't yet know what "quantization" means,
and you clearly don't know what "error" means either.
Your own posts admit that if there are only one-and
two-level variations in individual channels, there is no problem. You can
demonstrate to yourself that that's what you get--providing that what you
are converting is a photograph, not a helicopter, a computer-generated
graphic, a tangerine, or other items of similar relevance.
If you have a color photograph that demonstrates any
significant problem, let's have a look and analyze it. That is what Ric has
offered to do with his masking image, and that's the right way of
approaching it. I look forward to seeing what he has, as I would look
forward to seeing any real images that you have.
However, when it is so easy to test the thing that we
are actually interested in, there is a very limited utility for
"proofs" that depend on histograms, computer-generated graphics,
gamut charts, and statistical analysis by people who don't know what
standard deviation is. So, on behalf of the moderators, I ask that all
refrain from posting such materials in the future unless they can be backed
up with real images.
Dan Margulis
___________________________________________________________________________
Posted by: "Richard Wagner"
Sun Oct 8, 2006 8:43 pm (PST)
Re: “Quantization” and Rounding
“Error”
On Oct 8, 2006, at 10:30 AM, Dan Margulis wrote:
This is substantially the same statistically illiterate
argument that was
posted two weeks ago. In response, I asked if you were
able to show
a photograph that exhibited this damage.
Dan, it is actually your approach that is
"statistically illiterate." I asked you two weeks ago to explain,
specifically, how you arrived at your "statistical" conclusions,
and you never responded. I have never seen a post from you that contained a
valid statistical analysis, or a description of how to go about one. It is
no wonder that the color scientists refuse to engage in debate with you.
"Show me a picture" is not a statistical analysis. Coming up with
a "mean and standard deviation" without describing exactly what
you are doing is statistically meaningless. You apparently do not even
recognize the difference between numerical analysis and statistical
analysis, since you often misuse the latter term in place of the former.
The moderators have discussed the threads that took
place in my
absence last week (prior to this particular post) and
we all agreed
that there is no further room for "proofs" of
this nature when it
is easy to prove or disprove the point with real
photographs.
This again demonstrates your anti-scientific bias, but
here goes. Take any "real photograph" that is 8-bit RGB. Copy it.
Convert the copy to Lab, then back to RGB, and save it. Generate the
difference files between the two. The difference file will not be black, as
some of the pixels will have changed. If your original file was perfect to
begin with, this change is "damage." Mathematically, it is
"error." Take the same image, convert to 16-bit, and do the same
conversions, then convert back to 8-bit. The file will be closer to the
original file than was the 8-bit round-trip file. There is less
"damage" or "error." Whether or not the
"damage" is important in any particular image is a completely
different question. But it is a fact that 16-bit conversions to Lab produce
less encoding error than do 8-bit conversions. I used a file containing all
colors possible in 8-bit sRGB to show that it was not any specific color
that was a problem, and to show what the **maximum** error could be. Why?
Because that's how scientists work.
These conversions have been tested extensively. Nobody
can show any
damage with real images even under unbelievably extreme
conditions,
such as (with images specially chosen to be sensitive
to damage)
converting and reconverting 100 times, or converting
and
reconverting multiple times at different phases of the
correction.
This is a concrete demonstration that you do not
understand (or accept) the science underlying the color space conversions.
There is no difference between "real images" and "computer
generated images" as far as the CMM is concerned. Both images are
processed mathematically in exactly the same way. If a color changes in a
"computer-generated" image because of a color space conversion,
it will change in a "real image" as well. The only thing that can
exist in the images are numbers (0..255), and the CMM could care less where
they came from.
As I pointed out previously, "converting and
reconverting 100 times" is statistically meaningless in this case. All
of the changes occur on the first round of conversions to Lab and back, and
further conversions give exactly the same results. There are lies, damn
lies, and (misused) statistics, particularly by those who have no training
in statistics.
We have been hearing this "data loss" inanity
at least since 1994,
when one Photoshop expert announced that neither CMYK
nor LAB
should ever be used in color correction because the
conversion
inflicted "catastrophic damage" on the image.
His "proof"? The same
idiotic "levels loss" demonstration that we
still hear about today,
which demonstrates only that colorspaces have different
gammas.
I have never used the term "catastrophic
damage" nor have I implied anything close. The "damage" or
more correctly "error" is small, but real, and it is
mathematically demonstrable. What this does demonstrate is that there is an
inherent encoding loss in converting from RGB to Lab, and it is far more
pronounced in 8-bit than in 16 bit. It is not at all unexpected if one
understands encoding, mathematics, or even computer programming. If you
encode A into B, and B is smaller than A, you can't get always A back. In
this case, there are simply not enough numbers available in 8-bit Lab to
encode all 8-bit sRGB colors, so more than one sRGB color gets assigned the
same value in LAB - i.e., some colors get lost in the conversion.
Your "proof" is similar. You take for granted
that some normal
effect of conversion is wrong, and then because the
conversion
behaves normally, you announce that you have proven
that there's
damage.
I don't know what "normal effect of
conversion" you're talking about. Ideal image transformations are
mathematically invertible functions. You should be able to "convert
back" to get exactly what you start with.
This is just what you did two weeks ago--you invented
an irrelevant
statistic, "unique colors", and then
announced as a given that it indicated damage.
This is more of the same: you announce as a given that
any
variation in any channel is "damage" and
proceed from there. As a
result, the rest of the post is worthless.
You obviously have a very weak grasp of mathematics and
numerical analysis. I never stated that "variation in a channel is
damage." I have shown that conversions of an 8-bit image from RGB to
Lab will usually reduce the color information in the image. Whether this
has any practical implications for any given image depends on the
particulars of the image and person using the image. That is an entirely
different question that is not directly related to the previous scientific
statement of fact.
Statistical process control is tricky, and fools many
experts.
This has nothing to do with "statistical process
control." This has been a specific discussion about encoding. You have
still not defined what you are determining the "mean" and
"standard deviation" of with your "statistics." We are
not discussing variations in ink densities in a press run or some other
manufacturing process that is subject to "statistical process
control."
You, on the other hand, even after Lee Clawson was kind
enough to
link the definition, don't yet know what
"quantization" means, and
you clearly don't know what "error" means
either.
Your condescending remarks are without merit.
Your own posts admit that if there are only one-and
two-level variations in
individual channels, there is no problem.
As a practical matter, for most, if not all people on
this Listserv, one or two-level errors in each RGB channel will not be
significant. However, direct 8-bit conversions from sRGB to Lab may result
in "errors of up to 24 levels in the reds, 9 levels in the greens, and
10 levels in the blues." These are the mathematical upper limits,
found by looking at all possible colors for an sRGB conversion. These
errors may be significant in some images, to some people (or they may not).
Significance depends on context. Ric may have an example where the errors,
for him, are significant. The errors are often much different (larger) with
other RGB color spaces, but you would not know that without doing the
analysis.
If you have a color photograph that demonstrates any
significant problem, let's
have a look and analyze it. That is what Ric has
offered to do with
his masking image, and that's the right way of
approaching it.
I would like to see Ric's image of interest put through
the same comparisons as the test image I used. It would be very
informative.
However, when it is so easy to test the thing that we
are actually
interested in, there is a very limited utility for
"proofs" that
depend on histograms, computer-generated graphics,
gamut charts,
and statistical analysis by people who don't know what
standard
deviation is. So, on behalf of the moderators, I ask
that all
refrain from posting such materials in the future
unless they can
be backed up with real images.
This again demonstrates your significant bias against
real science. That's unfortunate, but not unexpected.
--Rich Wagner
___________________________________________________________________________
Posted by: "Iliah Borg"
Sun Oct 8, 2006 9:15 pm (PST)
Re: “Quantization” and Rounding
“Error”
Dear Richard,
Sunday, October 8, 2006, 8:35:34 PM, you wrote:
"Show me a picture" is not a statistical
analysis.
Of course. A good remark about statistics is attributed
to Disraeli.
"There are three kinds of lies: lies, damned lies
and statistics."
Floating point is even better then 15 bit, and allows
to sell more new computers. I can see how floating point can help raw
processing, but it is not about slight histogram variations. It is visible
and printable difference.
The irony is that higher precision already won the
battle, but in a different field. It is demosaicing, white balance,
sharpening, noise reduction - and as the method is already developed and
implemented in hardware, colour space conversions and curve adjustments
will follow. They are faster, close to instant in OpenGL.
If your original file was perfect to begin with
focus is perfect, lens is sharp and clean, no
vibration, etc...
Excuse me, but the damage you are speaking of is not
real life. It is about a spherical steed in vacuum.
--
Best regards,
Iliah Borg
___________________________________________________________________________
Re: “Quantization” and Rounding
“Error”
Posted by: "Olivier"
Mon Oct 9, 2006 4:48 am (PST)
--- In , Richard Wagner wrote:
Dear Richard,
Your will not to take for granted a cmd+click advise
without further explanation is legitimate. One might rightfully question
not the assertion itself since it has proved hard to challenge but the
reason of such an assertion which was the topic of some other posts.
Oppositely to you I believe it is very hard to
demonstrate the round trip is visibly degrading the image in an 8bits
workflow for what you need is to make the quantization error widely
multiplied.
We know that small gamut spaces need less bits to
encode colors (keep in mind the distinction between codes and colors since
"visibility" is at stake here) for the distance/variation between
two colors is less and thus you have enough of 8b. Again you would need to
think it twice to show on a real world image this effect. Yet I believe,
but don't assert, you could, under very specific conditions, but seldom in
the real world.
Looking at the Lab structure we know a and b chanels
code colors are based on opposition. In RGB, this is a bit different : any
R and G color could be the result of a combination of the two, thus you'd
have so many available code combinations that those two channels are not
good candidates to show up any potential issue in the round trip. Your're
left with the RGB B chanel : this single one encodes both blue and yellow.
We know that yellow will show in the dark part while blue in the light one.
Now comes the gamma factor, since RGB encodes both ligthness and color on
each chanel. We also know that high gamma will preserve shadows coding,
while with a low one, this is not the case. So here we have some first
conditions for a potentially troublesome color : saturated yellow in a
possible Apple color space. Yet this is still not enough : a bright yellow
alone might not show any reasonable issue, you would also need to be in an
area of the RGB space gamut where yellows codes could be the furthest
apart. This is where I am uncertain of the outcomes, because only wide
spaces can surely have this peculiarity.
Now let's consider a real world move on a real world
image. I may have one myself as an example (though I have to say I haven't
tried for I hate testing) : I live near Versailles (France) and the castle
has lots of outdoor guilded statues (hope this is the right word, I mean
covered with actual gold sheets) that have just been renovated, so those
show violently bright yellows. Under certain day lightening some parts will
be bright gold, other will be more in mid-shadows (yet still very saturated
yellow), I might want to give some additional pop to those later ones.
What I need is a brightening and some additionl
saturation of the gold mid-shadows while leaving the other golden parts as
they are. All this would eventually lead me to think I want to make a Lab
move. My move would be : convert to Lab then 1) lighten not enough yellow
yellows 2) saturate those not enough yellow yellows, then whenever ready
move back to RGB. I'd probably do without a mask.
Let's think about what will happen to the RGB B chanel
considering I have an low gamma work space. Saturating the not
enough-yellow yellows will darken the B chanel, yet lightening those will
oppositely lighten the B chanel and if bad luck is with me it will do so in
the dark part of the B chanel where I have not so many bit for I am in a
1.8 gamma space and where these two moves globally lighten RGB B dark
values. If you make this move what you will probably find in the B chanel
is that at some given B values you have created some noise and some
posterization of yellows. But as I said this is not enough for I doubt you
will see the difference on your screen. So you would probably need an
additional correction in RGB to have the bad B chanel really showing up : a
correction of any sort that would hit RGB B chanel is probably the thing
provided it further lightens it.
The other way you could do that is also as mentionned
in other posts with a Lab b mask, yet I don't find this is a good example
for I am not sure one would want to be in Lab for creating such a mask :
not only you may find noise but you also find a chanel that is not
"edgy" enough for you to use in some conditions and depending on
the move you want, so both combine, I feel you get a poor mask if you need
a precise masking. My feeling only here.
As I said I haven't tried that nor am I willing to. So
I might very well be totally wrong in all...and I am sure this will be
pointed out to me. Let's hope for courtesy.
Olivier Desmaison
___________________________________________________________________________
Posted by: "Mark Segal"
Mon Oct 9, 2006 9:16 pm (PST)
Subject: Re: [colortheory] Re: 'Quantization' and
Rounding 'Error'
I don't understand why it should be such a big deal in
this case, or why the mention of 16-bit causes such dramatic increases in
blood pressure and pulse to many on this
List.
Rich,
Neither do I. it should be completely uncontroversial,
because it is no big deal. People should work in whatever bit mode suits
their purposes. The same goes for working colour spaces. I think the
continuation of either of these issues is completely pointless.
Mark Segal
___________________________________________________________________________
Posted by: "Dan Margulis"
Tue Oct 10, 2006 5:27 am (PST)
Re: “Quantization” and Rounding
“Error”
Rich Wagner writes,
Take any "real photograph" that is 8-bit RGB.
Copy it. Convert the
copy to Lab, then back to RGB, and save it. Generate
the difference
files between the two. The difference file will not be
black, as
some of the pixels will have changed.
You're doing good so far.
If your original file was perfect to begin with, this
change is "damage."
Correct again. Similarly, if the population of Nome,
Alaska, is 40 million, it is the most populous city in the world.
A "perfect" original file is what a
computer-generated gradient is, not a photograph. This is one of the
several reasons that such gradients are worthless in evaluating how
photographs behave during colorspace conversions.
We do not convert computer-generated gradients into
LAB. We convert photographs. They come from cameras. Each camera's
performance varies considerably with, among other things, the lighting;
particulates in the air; how clean the lens is; how well-focused the shot
is; how much garbage is floating around inside the body, and how steady the
photographer's hand is. The sensor's performance varies with temperature,
humidity, and voltage fluctuation; and then the whole mess gets demosaiced
in a hamhanded fashion, whereupon it may get its bits blasted and blurred
to hide the inherent noise, and then some imperfect program chews it up and
regurgitates it into a readable format.
These, coupled with the considerable imperfections of
every output process, suggest that whatever values the camera returns are,
to be charitable, highly dubious. The whole problem with this thread is
that you have somehow taken it into your head that this odiferous swamp of
inaccuracy is so pure that even the slightest variation from it is an
"error."
As a practical matter, for most, if not all people on
this Listserv,
one or two-level errors in each RGB channel will not be
significant.
This is a wildly conservative statement. The 24-level
variations that you are claiming would definitely be significant, but the
lower limit is higher than two. As I pointed out to Mark, even with
occasional three-level variations the odds against a single print element
being visibly different are hundreds of billions to one. To have any hope
of seeing a difference you'd have to have a significant number of
four-level variations in the green and/or five-plus variations in the red
or blue. One-level variations are statistically meaningless and two-level
variations could only possibly be significant in the green. Nevertheless,
to put this nonsense behind us, let's agree that only one- and two-level
variations are acceptable. Since you can't or won't do the exercise with a
real picture, I'll do it for you, and since you don't know standard
deviation, I'll go the crude route with pixel count.
The results are substantially the same across all
photos that might go into LAB. I would suggest the first two out of my book
because, obviously, they feature the kinds of image that are most
advantageous for LAB and are taken there most frequently.
In each case, I have gone to the RGB source files,
assigned sRGB, converted a copy to LAB, reconverted to sRGB, and compared
the two as you describe, in the hopes of finding these 24-level differences
that you have promised us. The results of the pixel-by-pixel comparisons of
the two files (times two images) are as follows, remembering, please, that
green is more important than the other two put together:
Figure 1.1A GREEN:
Identical: 80.19% of pixels
Vary by one level: 19.81%
Vary by two or more levels: none
RED:
Identical: 51.38%
Vary by one level: 46.06%
Vary by two levels: 2.56%
Vary by three or more levels: 0.0004%
BLUE:
Identical: 59.77%
Vary by one level: 40.12%
Vary by two levels: .11%
Vary by three or more levels: none
Figure 1.2 GREEN:
Identical: 82.93%
Vary by one level: 17.07%
Vary by two or more levels: none
RED:
Identical: 56.72%
Vary by one level: 42.45%
Vary by two levels: .83%
Vary by three or more levels: none
BLUE:
Identical: 55.97%
Vary by one level: 43.68%
Vary by two levels: .35%
Vary by three or more levels: none
Two files do not have to be pixel-for-pixel identical
to be identical for any conceivable practical purpose. That's what these
numbers show with these two pairs. If you feel otherwise, you are cordially
invited to demonstrate how you could get a quality difference with any
rational procedure under any output conditions between any photograph that
went sRGB>LAB>sRGB as opposed to one with no such unnecessary
conversion.
The results, repeated over and over again, clearly show
that we *don't* get 24-level differences or anything else remotely close to
causing damage if we convert photographs. Demonstrations that purport to
show that we *would* get them if we converted something other than a
photograph into LAB are valueless.
What would be valuable is a demonstration that the
above results are not typical of what one gets when converting the type of
color photograph that people are likely to use LAB for. The only way to
show that is with a photograph. If you have such a photograph, let's see
it. If not, please move on to some more constructive topic.
Dan Margulis
___________________________________________________________________________
Posted by: "Bartlomiej Walczak"
Wed Oct 11, 2006 2:43 pm (PST)
Re:"Quantization" and Rounding
"Error"
Hello,
Just to chime in:
Dac> We do not convert computer-generated gradients
into LAB. We convert
Dac> photographs.
True but gradient can be a model that will tell us what
exactly happens in the conversion with each and every colour. While it is
not a real-world photograph, it can illustrate what one can expect from the
conversion. Because the conversion happens on a pixel-by-pixel basis, it is
the perfect model. It doesn't have to be a gradient. It can be a set of
randomly placed pixels. It's only important that all the colours are there,
because then we will see how each colour behaves in the conversion. And
that's all.
If two different colours are smashed into one during
conversion, then this is what we want to know.
And another thing - I am a movie editor. I often work
with computer generated images which are then composed with the
"real-world" footage. My job is to make them look as real as
possible. I rely heavily on the techniques found in your books (they are
awesome, btw). But if conversion to Lab makes my graphic a mess (and
gradient with banding in editing is considered a mess), then I want to know
if and how to avoid it. So please realize that there are not only
photographers on this list.
Dac> The whole problem with
Dac> this thread is that you have somehow taken it
into your head that
Dac> this odiferous swamp of inaccuracy is so pure
that even the
Dac> slightest variation from it is an
"error."
Yes, they are not "pure" (by this definition
nothing is but it is irrelevant). But why add to this process when you
don't have to?
All the best
Bart Walczak
___________________________________________________________________________
.
Posted by: "Dan Margulis"
Sat Oct 14, 2006 10:02 am (PST)
Re: Converting gradients
Bart writes,
If two different colours are smashed into one during
conversion, then
this is what we want to know.
Because I did not see this message before I posted my
request to end the "quantization" thread, and because it brings
up the new topic of converting gradients in live jobs, I will respond
briefly from the road.
Computer-generated gradients and photographs share
almost no characteristics. If two of 16 million different values become one
in converting a computer-generated graphic, there is a potential problem.
If this happens with a photograph it is of no significance, of no possible
relation to anything that may happen down the line--it is a completely
irrelevant factor.
And another thing - I am a movie editor. I often work
with computer
generated images which are then composed with the
"real-world"footage.
My job is to make them look as real as possible. I
relyheavily on the
techniques found in your books (they are awesome, btw).
But if
conversion to Lab makes my graphic a mess (and gradient
with banding in
editing is considered a mess), then I want to know if
and how to avoid it.
I agree that banded gradients are a mess. For this
reason, In all my books, I have pointed out that gradients should be
created in the destination colorspace when possible, because all forms of
conversion damage them. If you would like to verify it yourself, the only
correct procedure is to test it by converting computer-generated
gradients--not photographs--to and from LAB and other colorspaces. This
should make clear why the practice should be avoided.
Demonstrations with *photographs* that purport to
predict how computer-generated gradients will behave during conversion are
worthless. The converse is equally true.
Dan Margulis
Receive Schedule
and Pricing updates.
If you let us know how to contact you, we will enter you into our monthly drawing for a free 2-day class (A $895 value).
- Get free video training
- Get free tips, tricks and shortcuts
- Get special sale pricing and free seminar invitations
Sterling Ledet & Associates, Inc. 4100 Perimeter Park South Atlanta, GA 30341
work sales@ledet.com work administration@ledet.com work webmaster@ledet.com
P work 770.414.5007 P work 877.819.2665 F fax 312.896.9094
- Authorized Training Provider
Copyright 2012. Sterling Ledet & Associates. All Rights Reserved.
Design by Digital Scientists, a Conversion Optimization Company.
