[argyllcms] Re: Camera profiling with bracketing

  • From: Gerhard Fuernkranz <nospam456@xxxxxx>
  • To: argyllcms@xxxxxxxxxxxxx
  • Date: Mon, 14 Mar 2011 12:07:14 +0100

Am 13.03.2011 04:46, schrieb Iliah Borg:
 From manual sampling using RPP:

STDEV_R STDEV_G STDEV_B
0.36    0.31    0.32
0.34    0.22    0.27
...


Sampling with scanin

STDEV_R STDEV_G STDEV_B
0.58    0.64    0.66
0.56    0.59    0.61
...

Hi Iliah,

well, basically even a standard deviation of 0.6 does not really sound alarming, 
as the uncertainty of the mean of a sample of size N, drawn from a random 
variable, is even by a factor of sqrt(N) lower than than sample standard 
deviation. So if I assume patch ROIs of >= 100 pixels, then then uncertainty of 
the RGB_{R,G,B} numbers would be expected to be less than 0.6 / sqrt(100), i.e. 
less than 0.06, and that would be less than 0.1% of full-scale.

But of course this basic rule needs to be treated with caution here, because 
choosing different patch ROIs does not generate different samples from the same 
probability distribution, but also introduces different amounts of outliers of 
different magnituge.

So it would quite interesting, how much the RGB_{R,G,B} determined by manual 
sampling and scanin actually differ. Could you post both sets of RGB_{R,G,B} 
numbers?

And next it would be interesting, whether computing a robust mean of the pixels 
of a patch (e.g. trimmed mean, or even median) instead of simply averaging the 
pixel RGB numbers would significantly improve the match between the manual and 
the scanin method.

Regards,
Gerhard


Other related posts: