[argyllcms] Re: colprof -r parameter
- From: Gerhard Fürnkranz <nospam456@xxxxxx>
- To: argyllcms@xxxxxxxxxxxxx
- Date: Fri, 31 Jan 2014 01:54:37 +0100
Isn't there also a -rs option for suplying a relative scaling for the default
raw factor? 0.3, 0.1, 0.03, 0.01,... (i.e. geometric scale with 2 steps per
decade) are IMO good candidates to try for less smooth profiles than default.
The defaults seem to be indeed rather smooth, but an "optimal" value which e.g.
minimizes cross validation errors does not always give pleasing results on the
other hand (so it may be safer to renounce 1 or 2 dE for better smoothness).
--
Best Regards,
Gerhard
Graeme Gill <graeme@xxxxxxxxxxxxx> schrieb:
>BC Rider wrote:
>> Hi,
>>
>> When using the default value for the -r parameter I get (for
>example):
>>
>> Peak err = 2.348581, avg err = 0.554618, RMS = 0.655012
>>
>> And when changing the -r value to zero I get:
>>
>> Peak err = 2.185529, avg err = 0.518934, RMS = 0.615882
>>
>> Since -r=0 is a perfect, noiseless instrument, why don't the errors
>go to zero at those points?
>
>Hi,
> the -r parameter is translated through a table into the actual
>underlying
>smoothing values based on various scaling factors and experience, and
>the interpolation
>algorithm won't work with no smoothing, since it would then have no
>means of doing
>interpolation. In addition, if you have two points within the same
>locality as a grid
>node that are inconsistent with the grid interpolation (ie. think two
>different
>measurements for the same device values), then of course the profile
>can't satisfy both
>at the same time, hence a minimum fit error.
>
>You can set the raw underlying smoothing factor using "-r rX.XXXeX",
>but be warned
>that off hand I'm not sure what value you would start with (it could be
>as
>small as 1e-8, or as large as 1.0, I can't recall), and will vary with
>dimensionality
>and number of test points.
>
>Graeme Gill.
Other related posts:
