[argyllcms] Re: Camera profiling
- From: <graxx@xxxxxxxxxxxx>
- To: <argyllcms@xxxxxxxxxxxxx>
- Date: Sun, 7 Mar 2021 11:08:36 -0500
Yves,
Beyond using a good camera;
Beyond shooting the target perfectly;
Beyond the quality of the measurement data;
I'm convinced there is a limit to how good a matrix/LUT model can fit data
which does not conform to the Luther-Ives condition.
If that was the case, we'd be in a different 'world'.
In my humble research, I found relating RGB to XYZ through linear algebra
works "mostly".
I'm persuaded I have to accept the "best" mathematical models can do and
just go on with fixing what remains to be fixed, beyond excellent profiling.
About the "weighting", as far as argyll's profiles are concerned, this is
not the case, all colors are treated equally. But it's a good idea in
itself. There's no such thing as perfection.
About dynamic range, Yves, please someone correct me here but converting RGB
to XYZ through a simple matrix model does not involve any mapping?
I agree with you when you write :
I think, I'll try to create raw converted images that will provide a
"good" starting point and especially without any subjective transform
made by others and if in the end if they are not colorimetrically perfect
it will be of my own making.
You see, that's the kind of conclusion I'm slowly working towards.
ICC profiling gets us 90% of the way, the rest is up to us.
/ Roger
From: argyllcms-bounce@xxxxxxxxxxxxx <argyllcms-bounce@xxxxxxxxxxxxx> On
Behalf Of Yves Gauvreau
Sent: Sunday, March 7, 2021 8:01 AM
To: argyllcms@xxxxxxxxxxxxx
Subject: [argyllcms] Re: Camera profiling
Roger,
maybe you'll find what you need here:
https://ninedegreesbelow.com/photography/articles.html#profile-digital-camer
a
Generally speaking, the quality of the entrants to a process are directly
responsible for the outputs one gets. I'm not saying there is something
wrong with your data but from my own experiments,
I see it's not easy to have a good shot of a target, because this, because
of that, there is always something that make a shot less then ideal.
All this being said, I see much better result then this in some papers I've
seen but are these just over fitting results, I don't know.
With the idea of creating a "general" purpose profile, "weighting" some
color more or less then others, would be IMHO contrary to the idea of
creating a general purpose profile.
Lastly, most camera have a larger dynamic range (DR) then is possible with
most if not all output device out there so if you add this to the equation,
you have on top of colorimetric priorities, you have gamut mapping to
consider and we also have tone mapping to add to all this.
IMHO, it would seem converting a raw image to something pleasing to look at
on screen or print is inherently a non objective and most likely a non
colorimetric process as well, especially if you put any degrees of artistic
license to your editing process on top of that.
I believe most DNG Camera Profile (DCP) have a tone mapping curve embedded
into them, unless you make your own of course and this curve alone makes the
profile less then ideal colorimetrically speaking.
I think, I'll try to create raw converted images that will provide a "good"
starting point and especially without any subjective transform made by
others and if in the end if they are not colorimetrically perfect it will be
of my own making.
This objective will make creating either an ICC or DCP profile much easier I
would think.
~Yves
On 2021-03-06 4:32 p.m., graxx@xxxxxxxxxxxx <
mailto:graxx@xxxxxxxxxxxx>
wrote:
A few notes about recent camera profiling experiments using argyll's
excellent colprof, if I may.
Out of a simple ColorChecker 24 chart, under direct sunlight, with a Nikon
camera, I am getting :
With -as (general shaper curve and matrix profiles)
DE76 AVG = 2.36 MAX = 8.73
With -ax (cLUT base table profile using a PCS of XYZ)
DE76 AVG = 4.15 MAX = 11.89
I was "hoping" for "improvements" i.e. smaller DE76s, beyond argyll color
profiling, but, the more I read about profiling digital cameras in general
the more I understood that there is only so much one can expect.
Two observations.
One, some colors are "emphasized" under one regime over the other.
For example, the Yellow-Green patch is always "overstarted" under a shaper
matrix profile where it is closer to its reference value in a cLUT profile.
Grays were less neutral, too, under my cLUT attempts.
Also, some site suggested giving some color some "weighting" by explicitly
minimizing a dE2000 kind of difference?
I thought it was a neat "empirical" idea.
And as I explained to a friend, if there was one set of color I would
minimize, for general photography work, it would be skin tones.
And in argyll's matrix profile, the Light Skin patch comes out to 1.12
deltaE. So, I think that's excellent.
The other idea is that, at some point, I have the impression that we hit the
law of diminishing returns, as digital cameras inherently have color filters
that are not linear combination of the human color matching functions. So it
will always be an exercise in compromise.
I read on some site (darktable) that 24 samples was not sufficient for
driving some "spline" profiling models?
Suggesting, perhaps, that larger sample sets, such as 140 patches in the
form of the ColorChecker SG, for example, would result in "better
performance" out of the model?
I should experiment with SG profiles in argyll next, for curiosity. But I am
not sure I'll get lower DeltaEs necessarily.
Best and kudo, as usual, Graeme, to you / Roger
Other related posts:
