[argyllcms] Re: problems hitting white point (color temperature, RGB gain/contrast) while maintaining brightness: dE not stable

  • From: Graeme Gill <graeme@xxxxxxxxxxxxx>
  • To: argyllcms@xxxxxxxxxxxxx
  • Date: Wed, 16 Dec 2009 09:44:54 +1100

David Heinrich wrote:

I followed Martin Weberg's suggestions. My readings still do not stabilize,
and are still bouncing around from a dE of 0.0 to even ~2 when adjusting RGB
gain channels.

This doesn't sound completely unreasonable, although one would always
like better repeatability and stability than this. Certainly the cheap
colorimeters show their limitations a bit when it comes to repeatability,
and displays similarly jump about a bit, depending on their technology
and design.

sudo ./dispcal -v -d1 -yc -t6500 -T6500 -P0.5,0.5,2.5 -o sony-12-15-2009

Hmm. I'm not sure what you're expecting with "-t6500 -T6500" - it needs
to be one or the other (Daylight or Black body). They are only subtly different,
and typically people will go for Daylight.

CRT, etc. What should I do from here? I presume if I go on to characterize,
the characterization will be hampered by these unstable values. Can I
characterize, or will it be worthless? (or maybe do several
characterizations and average them?).

No characterization is perfect. Even with good color devices and instruments
there will be repeatability variations. Simply picking up a spectrometer
and putting it down again on the same patch can give you a change in reading
of 0.5 dE or more. If one were to go to a lot of trouble you could measure it
lots and lots of times and get a more accurate characterization on average,
but is it worth it ? There are so many other sources of errors - device drift,
model fit, observer variation, instrument calibration, source profile errors,
and practical samplings of a color device are relatively sparse anyway, simply
because of the size of the gamuts (ie. attempting to sample an RGB gamut with
1 dE spacing would take about a million readings.)

I do remember from before it did say that when initially trying
characterization in several steps (using the dispcal output as a starting
point), colprof had "average errors of average errors of 5 or less, and
maximum errors of 15 or less" (this is what the colprof documentation
suggests is acceptable).

These are the model fit (or self fit) errors. They indicate how self consistent
the data set is, as well as how well (or badly!) the model fits the device.
Something like a matrix profile has a much stronger sense of model that cLUT 
based
profiles. The latter is constrained only by its smoothness criteria and
grid resolution. Looking at the these errors is merely a sanity check. Typically
if you've got samples swapped it will show up here.

Graeme Gill.


Other related posts: