[argyllcms] Re: targen options

Gerhard Fuernkranz wrote:
> Thanks for the info. Sounds tricky :-) So I guess, it places more dots 
> at locations with a tight curvature of the device->Lab mapping, and 
> fewer points at locations where the mapping is flat, does it? Do you 
> already have some evidence, whether the resulting profiles are "better"?

I haven't done the test runs to confirm whether it is actually better or not.
I was hoping to get the itteratively improved adaptive working, but
haven't managed to finish it.

One of the things I discovered in testing the non-adaptive routines,
was that the result was rather sensitive to various factors, such
as the distribution of the verification values, the process used to
create the reference device response, the reference device responses
etc. Overall the distribution of test points seems to make a small
but just measurable difference to the profile accuracy (at least
for the non-adaptive approaches).

My previous algorithms were skewed due to the use of non-random verification
values (ie. there is a correlation between the verification distribution
and the test distribution, so you appear to get better results if the algorithm
you are testing has a similar distribution to the verification points.)

One thing that was noticeable was that for particular device characteristics,
you can just be unlucky in your sampling. If the algorithm hits a particular
spot (presumably one of high curvature), you will get an OK maximum error. If
you don't hit that spot, your maximum error might be twice as high.
The average errors were more consistent.

The above observation was what motivated developing the adaptive approach.
The algorithm itself was inspired by the following two papers:

"The Farthest Point Strategy for Progressive Image Sampling",
Yuval Elder, Michael Lindenbaum, Moshe Porat & Yehoshua Y. Zeevi,
<http://www.cs.technion.ac.il/users/wwwb/cgi-bin/tr-get.cgi/1996/CIS/CIS9613.pdf>

"Fast Marching farthest point sampling",
Carsten Moenning and Neil A. Dodgson,
<http://www.cl.cam.ac.uk/users/cm230/docs/pdfs/FastFPS.pdf>

The other notable result that (realized in retrospect from peoples reactions)
was that evenly distributing in perceptual space gave consistently worse 
results.

Graeme Gill.

Other related posts: