[argyllcms] Re: RGB Printer profiling and ColorSavvy CM2C
- From: Graeme Gill <graeme@xxxxxxxxxxxxx>
- To: argyllcms@xxxxxxxxxxxxx
- Date: Tue, 02 Aug 2005 11:41:14 +1000
Gerhard Fuernkranz wrote:
Counter intuitively, my experimentation with perceptually evenly
distributed
test points was that they produce worse results than evenly distributed
Oh really! And I was always blindly assuming, that the perceptually
distributed points, and targets tailored to the particular device
characteristics, are one of the great, unique features of Argyll :-) But
obviously you are right. Having never tested this before and being
curious, I've now also done a brief test: Used an existing CMYK profile
as reference, created 200 patches with and without "-c", performed
measurements with fakeread and created the profiles. Then I created
10000 test patches (independent of the the patches used for profile
creation), performed measurements with fakeread and checked the profiles
with profcheck against the 10000 test points. The profcheck results of
both profiles were very similar, the profile generated from the
perceptual target was marginally less accurate (but I would not call it
a statistically significant difference).
I did a bunch of testing for the paper I wrote last year for the
CIC. To test well, you need to choose a very large, uncorrelated test
set. I ended up using 10000 or more random test points to check the
profile fit. The factor that affects the peak error the most is whether
the profile set happens to land a test point at a critical location
or not, which is kind of random. The best results were generally
with what is the default "full spread" algorithm.
So I guess we can conclude that 2-pass profiling (first creating a
preliminary profile, and then generating the final profile using an
optimized target) is rather waste of time and consumables, right? (and I
must admit, that's a different insight from what I expected previously)
In previous releases of Argyll there was an attempt at 2 pass profiling
using "perceptually uniform" distribution. There were two problems though.
One is that perceptually uniform doesn't improve things (as mentioned above),
and the other was the the algorithm intended to create even point spacings
wasn't doing it well enough.
The current release of Argyll does something quite different if the -A flag
is set to a non-zero value (and a profile is supplied to -c). It doesn't
try and place test points perceptually uniformly, it tries to place them
at points that most reduce the maximum resulting profile error.
I haven't tested this approaches effectiveness, and it could well
perform better than the current default approach. (Feel free to test it ! :-)
Graeme Gill.
Regards,
Gerhard
Other related posts:
