[argyllcms] Re: Possible whitepoint problem

  • From: Gerhard Fuernkranz <gfuer@xxxxxx>
  • To: argyllcms@xxxxxxxxxxxxx
  • Date: Tue, 11 Sep 2007 01:11:00 +0200

John Weissberg wrote:

> I have always been told that our linearization stage is trying to
> linearize K vs. "Optical density". I would have thought that density
> would correspond closely to L. Here that relationship is clearly not
> linear.

Density versus L* isn't linear either (see attachment), but has a much
lower curvature than the K vs. L* curve derived from your printer data
set. I'm afraid it is up to you to find out what did went wrong with
your linearization, I don't think that I can help here.

> Even though your method of adding 16 times the white point into the
> ti3 does force a good value of wtpt into the icc profile, I still
> believe there is some sort of problem with argyll's method of
> calculating the whitepoint.
> By comparison, when I used the exact same dataset (with only 1
> whitepoint in the dataset) to produce a profile with Gretag
> ProfileMaker  that ICC profile's white point had an L value of 97.

I guess PM assumes a different error model, i.e. it seems to assume that
the given measurement for the white patch is complety error-free, and
thus it attempts to fit this data point exactly, while Argyll basically
assumes that all patches may suffer equally from printer repeatability
and measurement errors, including the white patch.

The assumption that the measurements of the white patch are error-free
may be convenient (as it will give a result which you likely expect),
but it isn't perfectly realistic either - even measurements for paper
white are not reproducible exactly. On the other hand, the Argyll
assumption of homoscedastic errors isn't granted in practice too, since
the printer repeatability is usually better at/near paper white than for
colors with a larger ink coverage (and the presence of output tables
distorts the error distribution too). So the "truth" is likely somewhere
in between.

But anyway, your major problem is that your data set does not fit a
smooth model well, particularly not in the bright regions near white.
Btw, in the mean time I have even doubts, whether this magnitude of
error can be explained by the linearization and the printer
repeatability issues only. Are you absolutely sure that the measurements
are all correct? Are you sure, that the printer or RIP don't behave strange?

> In addition, when I use a gamut viewer to compare the gamuts of the
> profile made with argyll and that made with ProfileMaker, the profile
> made with argyll is completely missing a volume between L=89 and L=96.
> I think this missing gamut is related to the whitepoint problem.

This is not surprising for the absolute intent gamut, since the gamut
cannot extend beyond the assumed white point. If you display the
relative colorimetric gamut (iccgamut -ir), then both ones are expected
to extend exactly to [100,0,0].

PNG image

Other related posts: