[argyllcms] Re: CMP Digital Target 3 (not 003)

  • From: Graeme Gill <graeme@xxxxxxxxxxxxx>
  • To: argyllcms@xxxxxxxxxxxxx
  • Date: Mon, 09 Mar 2009 14:49:02 +1100

Gerhard Fuernkranz wrote:
(hopefully well-know) gamma to the converted raw images. Therefore if
one deals with camera raw images then it may IMO even make sense to
assume fixed shapers [preferably gamma 1.0, as captured by the sensor,
or otherwise the well-know gamma value which was explicitly applied by
the used raw converter] and to estimate the matrix only [Graeme, I guess
it should not be too hard to add an option to colprof for prescribing a
fixed gamma for the shapers, instead of estimating them along with the

Hi Gerhard, Pascal and others,
        of course anything is possible, but I'm not convinced yet of
it being desirable. While one can make lots of suppositions about
how particular devices work, some experimental proof that the
suppositions are correct would be good before changing how things
work on the basis of them.

One thing to keep in mind is that the profile isn't necessarily meant
to mimic the way a device works, it is meant to model the devices
overall behavior using the mechanisms the ICC profile makes available.
So the per channel curves will likely be different for best match
for (say) an XYZ PCS CLUT profile than an L*a*b* PC CLUT profile, etc.

The other aspect is that I'm not happy with trying to second guess the
modeling. While in particular cased one may be able to make guesses
about how one model will better fit than another, having even finer grained
options seems to me to be something that most people won't understand
or cope with (apart from randomly changing things to see if they
work better), and misses what I see as the real issue:

Why are the models not as accurate (subjectively) when they
are more accurate numerically ? I'd rather fix this so that
you don't have to make assumptions about the device or know
to set intricate options to get the best possible result !

Some examples of such a situation would help (ie. the .ti3 data
plus the colprof params and example images for such a situation).

response to a _fixed_ spectrum at varying intensity). The deviation of a
linear sensor from a true linear response is likely smaller than the
error of the response curves estimated in this way. Thus simply assuming
gamma 1.0 for a sensor which is known to be linear may be more
appropriate than trying to estimate its response.]

Right, but another interpretation is that if this is the case, the smoothing
is inadequate (too much noise is making its way into the model), and imposing
a linear curve forgoes the possibility of modeling any other sources
of nonlinearity in the whole system.

Graeme Gill.

Other related posts: