[argyllcms] Re: Amount of patches

Pascal de Bruijn wrote:

So my basic question would be, is there a "logical" "sensible" amount
of patches to read for creating a matrix profile?


Generally fewer patches are needed for the more restricted behavior
model based type profiles, such as gamma/matrix or shaper/matrix.

At its most basic, it comes down to the number of parameters a model has.
Ideally you need at least as many measurements (numbers) as there are free
parameters to the model. For gamma/matrix that's something
like 12 parameters, so a small number of patches may be sufficient.

But in reality, the readings are not noise free (not perfectly
accurate), and you may want the simple model to represent the
average of the device behavior over its gamut, so you need to sample
the devices behavior over its whole range of values to get an average
and reduce the influence of inaccuracy in each measurement. Past
a certain point though, the differences will get less and less
(the point of diminishing returns). This number will depend
on the profile model chosen and the behavior of the device.

The cLUT based models have many more parameters because they
divide the model up into small regions within the colorspace.
Typically there are too few patches for cLUT based profiles to
get any averaging effect, which is why the profile making process has
to apply some level of filtering (or regularization).

Graeme Gill.

Other related posts: