[argyllcms] Re: perceptual black too light

Gerhard Fuernkranz wrote:

Von: Graeme Gill <graeme@xxxxxxxxxxxxx>
But as far as I can tell, this doesn't differ from relative
colorimetric intent.


If you consider the monitor profile only, it does not differ
from relative colorimetric (at least if you assume that the
monitor observer is adapted to monitor white).

But if you consider the transformation from printer profile to a
monitor profile, then ICC-absoulte is indeed different from relative
colorimetric. Relative colorimetric maps paper white to monitor
white, but ICC-absolute maps the illuminant color (i.e. the color
of an illuminated perfect diffuse reflector) to monitor white, such
that for instance a blue-ish, FWA-rich paper is also simulated with
a slightly blue-ish tint on the monitor.

Furthermore, since paper usually has a reflectance of < 100%, the
luminance of paper white, after transformation to the monitor space,
becomes implicitly lower than the monitor white luminance, which also
helps to reduce out-of-gamut probles if tinted paper white is simulated
on the monitor).


Right, but it doesn't work properly the other way around. If I want
to simulate the absolute appearance of the display on some paper,
then using the ICCV4 defined intents, there is no way to do so.

Because an emissive display doesn't have a separate illuminant and
media, two distinct intents that can be applied to reflective display
have to be combined. The ICCV4 change, does this in a manner that
reduces functionality, leaving only 3 intents for emissive
displays, rather than the 4 of previous ICC specifications.
And it's all completely unnecessary. The same effect could
have been achieved by simply linking your paper source
in absolute, to the display in relative, if that's the effect
you wanted.

If src and dst illuminant match, then ICC-absolute becomes identical
to the "traditional" interpretation of absolute colorimetric. For
instance in case of a hardcopy proof, where the proof and the final
print are viewed under the same illuminant, it does not make any
difference whether the illuminant relative ICC-absolute intent or
the "traditional" absolute colorimetric intent is used for
computing the proof - both will give the same result.


I'm not interpreting the ICCV4 spec change to be a shift from
assumed media adaptation to illuminant adaptation. On the contrary,
the continued use of the media white as the point of reference for
"Relative Colorimetric" indicates to me the opposite. The transform
from measured (reflectance) absolute values to relative colorimetric
is still being done using the "wrong Von Kries" for no apparent
technical reason, but the media is still assumed to be adapted
white. The shift from using the emissive white point as the media
white rather to the illumination, has made side by side hard copy
to display proofing and spot color reproduction harder to achieve,
and injected uncertainty into the interpretation of display profiles.

But this seems to be the trend in a lot of the changes to the ICC spec. :-(
Changes made with the claim of "improving" interoperability, mostly seem
to be having the opposite effect.  The change to the PCS L*a*b* encoding
is a case in point. What was a very simple and unambiguous specification
in the ICCV2 spec., is now highly complex, and I have reason to suspect
that there are no CMM's out there that are compliant with the V4
specifications in this respect.

I'm not sure, what I should believe. On the one hand I tend to agree
with your newspaper example, on the other hand, if I'm viewing a book,
magazine of a sheet of paper in a usual everyday viewing environment
(living room, or maybe also outdoors), i.e. in an environment,
where the sheet of paper is yet another object in the scene, then
I think it is possible to tell whether the paper is rather blue-ish
or yellow-ish, even if neither the light source nor a reference
object with a known white color is visible in the field of view.
So there indeed seem to exist congnitive mechanisms in our vision,
which are able to estimate and discount the illuminant.


I'm not sure that's true. If you make a conscious effort to compare
the book to other things in the environment, then yes, I'm sure
you can make a judgement about it's paper color, relative
to the appearance of other white references you can see. But
if you are actually reading the book, I don't think you would
interpret it as being other than "white".

Eventually I guess, depending on the actual viewing situation,
the "truth" is likely somewhere in the middle, sometimes closer
to the illuminant color, and in other situations closer to
media white.


It still seems to be a subject for research papers, but it certainly
seems reasonable to assume that in many real world situations, the viewer
is only partially adapted to a particular white.

(Actually it is even more complicated, since particularly for
larger images, there is not a fixed adaptation state of the
observer, but the observer continuosly readapts, when the
eyes move through the scene (or image) ...)


I tend to assume that the same mechanism that is responsible
for white adaptation, is also responsible for things like
simultaneous contrast, and that retina receptors are adapting on
a very localized level. There do seem to be higher level, cognitive
effects involved as well though.

Regarding your newspaper example, are you sure, that the photos
have been rendered with a perceptual gray axis with paper chromaticity,
or are they probably rendered with a perceptual gray axis with D50
chromaticity, which just bends towards paper white (or actually
paper green) at the upper end?


I'm 99.9% sure, since the newspaper involved uses the RIPs for which
I wrote the color system, and Argyll was used to generate the profiles
for the proofing. The proofing has been done both by printing onto
the actual paper stock using an inkjet, and also emulating the
paper color using a color copier. The effect is purely subjective.
If you flip an edge of the paper over to lie beside the highlight
in a photo, it is clear that they are the same color. It's interesting
that broader areas of the paper do not look so white. Simultaneous
contrast helps accentuate the appearance of white in the photographs.

Btw:

In papers regarding research of chromagenic color constancy
(i.e illuminant estimation) algorithms, Morovic, Finlayson
and Hordley write

"The central area of the human retina - called the fovea - also
uses a filter, the macular pigment. This is the reason why there
are two standard colorimetric observers - 10° and 2°, depending
on the field of vision.


I was under the impression that the cones and rods had different
spectral sensitivities, hence the difference in 10 and 2 degree
sensitivities. Now it might be a matter of semantics as to whether
macular pigment is lumped into the response of the cones or not,
but given the different construction of the rods, I would be surprised
if they had the same spectral sensitivities to the cones, even without
the macular pigment filter. Certainly at the end of the day, our
spectral sensitivity for scotopic and photopic vision is
different.

It turns out that the 10° and 2° observer spectral sensitivities
are a chromagenic pair - simulating the human visual system and
using these two observers for the unfiltered RGB and filtered
counterpart, we achieve good illuminant estimation (significantly
better than using other methods)"

See for instance
http://photo.csail.mit.edu/posters/chromagenic.pdf

They don't claim or proof that the illuminat estimation of the
vision actuall does work in this way, but obviously the did a
computational proof of concept, that it could work in this way.


Yes, I remember this was presented at the last CIC.

With ICC-absolute intent, such an absolute match (for a side-by-
side comparison) can only be achieved with a D50 calibrated monitor
(or in the general case, with a monitor whose WP is calibrated to
the illuminant color used to view the print).


I think it can be achieved without this, if the application uses full
screen display, and if the CMM is smart enough to avoid clipping
the white. There would be more flexibility using this sort of approach.

OK, so you do NOT really split the src -> dst mapping into a
mapping from the (possibly skewed) src gray axis to the non-skewed
PCS [0,0,0]..[100,0,0] gray axis and then from PCS to the (again
possibly skewed) dst gray axis (where only the 2nd part gets
recorded in the B2A0 table), but you rather treat the src profile's
gamut more ore less directly as "PCS gamut", for establishing the
PCS -> dst device gamut mapping.


The difference between the old scheme and the new scheme,
is that the old scheme aligned both the white and black points
of source to destination, while the new scheme aligns only
the white point (the black point being modified only
as a consequence of the white point change.)

Yes that's what I think as well. IMO one needs very accurate
measurements and a very good printer repeatability, in order that
the default smoothness would be appropriate. But since this is
now tunable via command line arguments, this is no longer a big
problem - everybody can now choose his desired smoothness factor.


I think the default uncertainty assumed is a bit too small, and I
modified the optimal smoothing values the simulation arrived at
quite a lot (reduced it by 10x) when I verified against some real
world measurement sets. The 0.53 values gave minimal RMS errors
of the reference set to profiles (made from a sub-set), and restoring
the smoothing values arrived at by simulation, increases the
fitting error somewhat, but greatly improves the smoothness. A
further 10x increase in smoothing factors causes a more dramatic
increase in fitting error, so I seem to be at a "knee".

I don't remember, which numbers I did use exactly for which data
set. The repeatability of my laser printer is not too good (about
2dE average (or maybe even worse), for a set of CMYK colors
printed several times at different locations on the _same_ page).
So particularly for this printer, the default smoothness is IMO
far too low. But my laser printer is likely not a good general
reference, I think typical inkjets usually behave better.


I haven't yet done this exercise on my inkjet printer.

I've seen, you've added a new option to fakeread to add noise
to the measurements. But it looks like you add uniform random
numbers. I think that gaussian noise would be more appropriate.
I do not want to claim that real measurement and reproducibility
noise is perfectly gaussian, but it has a rather bell-shaped
PDF, not a rectangle. IMO the disadvantage of uniform noise is,
that the max. deviation is strictly limited, while a bell-shaped
PDF also can produce larger outliers (though with correspondingly
low probability), which I think is more realistic.


I wasn't really aiming for a perfect simulation, and the uniform
randomness made some things a little easier to calculate. The
next level of refinement would be to try and characterize real
world error distributions (ie. are they uniform and gaussian
when measured in XYZ ?), and to taylor the smoothing factors
appropriately in the L*a*b* colorspace (ie. translate it into
sample point weightings).

There seems to be a fair degree of latitude in the smoothing
values though, so I'm not sure in the end it would greatly
improve things.

Graeme Gill.

Other related posts: