Gerhard Fuernkranz wrote:
Gerhard, thanks for sharing your results.
I had not done the same exercise in XYZ space, but I'm afraid, the errors are heteroscedastic in XYZ as well. I would really be surprised if the repeatability error of a printer would be homoscedastic in XYZ space.
Yes, I'd imagine that the errors introduced by the printer won't be so well correlated to linear light, although I would expect the errors from the instrument to be. As a first order approximation, I would guess that the printer would generate a normally distributed error in the amount of ink it lays down, and then a (rough) model of ink level to color would translate the statistics into color terms.
I'm guessing that instrument errors will be most significant at the dark end, where slight errors in light level are magnified by their ratio to 0, and that printer errors might be most significant at the light end, where slight errors in the ink level, are magnified by their ratio to white.
since your rspl weights are weights for the suared error. As you do not support separate weights for each RSPL output dimension, but only a single weight per point, a reasonable compromise is likely to use 1/(variance_L+variance_a+variance_b) as point weight.
Actually, I think the latest code does have support (at the fundamental level) for different smoothness weights for each (input) dimension, and it would be easy to add different weights for the output dimensions, since each output channel is fitted independently (in fact, I was tempted to do this when I was doing the tuning, since the simulations indicated that the optimal L* value smoothness is different to the a* and b*, which is hardly a surprise.)
Given I was struggling to tame the complexities of determining good smoothing factors for combinations of different dimensions, number of sample points, and uncertainty level, I decided not to further complicate things at that time.
Graeme Gill.