[argyllcms] Re: Perceptual spread in targen for non-linearized devices

Nikolay Pokhilchenko wrote:

Thank You for the answers, Graeme! I'm agree with You that there will be no 
difference
between L*a*b and Jab in resulting profile. But I disagree yet that it's hard 
enough
seeing a reproducable trend between device space and perceptual space. Striking 
exampe
is non-linearized device. Take a look at attached files. There are two targen 
charts,
one is OFPS device-space spread (default targen -d3), other is perceptual space 
random
spread (targen -d3 -c"pre-prof.icm" -A0.93 -R). 

Note that -A0.93 isn't doing anything here, since you overridden the
default Optimised Farthest Point Sampling (OFPS) with Random.

There's no doubt it make a difference to the look of the test charts,
but it's measuring the resulting changes in the actual profile accuracy
that is the challenge.


I'm insisting upon that the perceptual
test patches spread (see target_perceptual.wrl) will be more informative for 
resulting
profile, rather the device spread (target_device.wrl). IMO the human vision 
sensation
is more important than the carefull device space charactrerisation. 

That is the intuitive conclusion. I did some extensive testing though, that 
seemed
to contradict intuition. It's quite hard to test, because you need a perfect 
underlying
device model to use an absolute reference, and the shape of this and the 
distribution
of verification points has a great influence over the results. For instance, 
using
perceptually distributed verification points makes perceptually distributed
test charts look good, and the same applies for device space distributed
verification points and test charts. Add to this the sampling noise inherent
in the situation, and it's hard to draw conclusions.

None the less, my testing indicated that (counter-intuitively), a device
spaced based distribution have a more accurate result than perceptually
based distribution, with OFPS being the best distribution, with this
trend being visible through the bias of the verification point distribution
choice.

The only explanation I can give is that the device characteristic is
a mapping between device and perceptual (PCS) space, so it's the
way these spaces are connected by the profile curves/surfaces that
is key, and that the spacing should not be too sparse in either
the domain or range.

I suspect that adaptive OFPS would be better if it was optimized,
but unfortunately I have been unable to find an algorithm that will
work for this, so you get a partially optimized results (ie. worst case
2:1 spacing error compared to optimized) when you use OFPS with -A
with the current software.


I'm prefering
nonuniform patches spread in device space (at input) in favor of uniform spread 
in
output space, leading, I hope, to less error at output.


I understand, but see above.

My results were published here:
<https://www.imaging.org/store/epub.cfm?action=showabstr&abstrid=32145>

Graeme Gill.

Other related posts: