[argyllcms] Re: Profile -r option

Hi Gerhard,

OK I understand, and I see the difficulty in making a call on this.

In this case the target scan is from a 35mm IT8 transmissive target scanned at 1600dpi. So the end result is that each patch on the IT8 has about 50x50 useful pixels in it. I'm not sure what scanin's sampling algorithm is, i.e. every pixel or just intermediate ones. If it samples and averages all of them, then that's 2500 samples per patch, so I would expect that noise and related imperfections in the patch would would have been virtually eliminated. (The scanner does multisampling to get rid of CCD noise too).

My scanner is fitted with a 'Scanhancer' lamp diffuser, which quite possibly introduces some spectral characteristics that don't follow a standard polynomial or gamma-curve TRC. That's why I would have tended toward a tighter fit LUT profile than a nice smooth polynomial curve.

Of course, there are also going to be measurement errors in the original target reference data, but I don't have a feel for how large those are.

Reproduceability: well, I would be disappointed in the scanner if it was introducing errors of up to dE of 4 from one scan to the next of the same slide. I really think the scanner is much better than that.

So my gut feel would be to run with the lower -r, perhaps 0.5.

Cheers...




Gerhard Fuernkranz wrote:
Milton Taylor wrote: 
Hi Graeme,

I'm interested in finding out more about the -r option. I'm remaking my film scanner profile, and this switch seems to have quite a dramatic effect on the results.
The parameter controls the trade-off between smoothness and fitting accuracy.
Leaving it off, i.e. default value, gives the best results in terms of avg and peak dE. Going from -r 0.5 to -r 0.75 and the error doubles. At -r 1.0 it's even worse.

You indicated that the default was too low, so why select a higher -r value if it gives a profile that is apparently not as good?

Just for info, in my case, -qh gave these results:
r = 0.25  avg dE=0.38 peak 2.8
r = 0.50  avg dE=0.48 peak 3.7
r = 0.75 avg dE=1.0 peak 7.8
r = 1.0 avg dE=1.6 peak 9.5

Which is the 'right' value of -r to use?
This depends mostly on your particular device and your particular measurements. If your device and your measurements were perfectly noise-free and perfectly accurate, and if the reproducibility of your device would be perfect, then indeed rather a low r (resulting in a low dE fitting error) would be desired.

However, when fitting noisy data (which are "contaminated" with random measurement errors, device reproducibility errors, etc), then an allegedly good fit (low dE) may be in fact not as good as it seems, and just fit the noise in the measurements, while a somewhat smoother profile with a larger dE may indeed predict the actual average device behaviour more accurately.

Since typical real-world devices and measurements are never perfect, your r=0.75 (or maybe even the 1.0) result looks more plausible to me than your r=0.25 result (which I'd rather expect to overfit the data).

On the other hand, if you would create a profile from a ti3 file which has been generated by averaging 16 (independent) measurement sets, then for this profile I would rather expect a fitting error like you got it with r=0.25, since averaging 16 independent measurements reduces random noise by a factor of 4 = sqrt(16).

Regards,
Gerhard

Cheers,
Milt



0.75 (or maybe even the 1.0) result looks more plausible to me
than your r=0.25 result (which I'd rather expect to overfit the data).
 <br>
 <br>
On the other hand, if you would create a profile from a ti3 file which
has been generated by averaging 16 (independent) measurement sets, then
for this profile I would rather expect a fitting error like you got it
with r=0.25, since averaging 16 independent measurements reduces random
noise by a factor of 4 = sqrt(16).
 <br>
 <br>
Regards,
 <br>
Gerhard
 <br>
 <blockquote type="cite">Cheers,
   <br>
Milt
   <br>
 </blockquote>
 <br>
 <br>
</blockquote>
</body>
</html>

Other related posts: