Am 06.02.2013 12:37, schrieb Graeme Gill:
If the underlying device response is additive then a 1D curves + matrix function will be more accurate in regions of the response that are not close to points in the measurement sample set. If on the other hand the device response is not additive, then a cLUT has the potential to have lower gross errors in those areas, but it will only be more accurate if there is sufficient density of measurement samples.
Hi Graeme, this is however not a property of CLUT profiles per se, but rather an implication of the non-parametric spline model used by Argyll to build the CLUT. For instance, smoothing spline regularization "pulls" the splines towards a linear function like Lab=RGB*M (in case of CIELAB PCS) which certainly cannot model the device behavior very well. Sure, that's the way how smoothing splines work... I was in fact already wondering whether it might be possible to have a regularization which does pull the splines towards a given simpler device model instead (e.g. matrix/TRC), so that high smoothing factors result in a CLUT which is (almost) equivalent to the matrix/TRC model, while low smoothing factors give the spine model more flexibility to adapt to the training data. The underlying idea is of course, that insufficient number of readings should not lead to an overall bad profile, but rather to a profile which is closer to the simple matrix/TRC model, and which deviates from the matrix/TRC model particularly in those regions which are sufficiently covered by measurements. Alternatively to a non-parametric model, one could of course also build a CLUT by resampling a parametric model (for instance the display model proposed by G.Sharma, having 9 TRC curves (3 for each RGB channel) whose outputs are then combined to XYZ by a 9x3 matrix, or what model ever). Best Regards, Gerhard