[argyllcms] Re: Best (smallest) profiling patch set

Graeme Gill wrote:
Interesting observations. The MPP fitter hasn't really been tuned up in regard to its behaviour with noisy readings. Probably the place to start, would be in increasing the four #defines TRANS_BASE, TRANS_HBASE, SHAPE_PMW and COMB_PMW, and/or altering the weighting scheme that TRANS_BASE and TRANS_HBASE participate in (xicc/mpp.c)

In theory the MPP model is better when some fundamental values are missing, such as only having the primary colorant values, and white. It will compute approximate secondary overlap colors, and use them as anchors for the resulting profile. The rspl behaviour under such circumstances might not be very realistic.
I justs looked closer at the .mpp file, and obviously the -qm MPP model has 64 parameters (per output dimension). So if there isn't some kind of regularization implied in the model which reduces the effective numbers of parameters, then the problem is indeed underdetermind with only 50 training data points. And in fact the -ql MPP profile with only 28 parameters seems to work a little bit better:

-qm, validataion agains test set: avg err = 1.423371, max err = 7.912226
-ql, validataion agains test set: avg err = 1.275329, max err = 7.226500
(both noise-free)


And here again the same comparison as yesterday, with mpprof -ql:

50 noise-free training patches, verified against 50000 noise-free test patches:

ICC: peak = 8.002047, avg = 0.941793, rms = 1.237705
MPP: avg err = 1.275329, max err = 7.226500
ICC via MPP: peak = 5.849207, avg = 1.169977, rms = 1.366576

50 noisy (-R 1.0) training patches, verified against 50000 noise-free test patches:

ICC: peak = 16.264535, avg = 3.130956, rms = 3.563022
MPP: avg err = 2.806883, max err = 15.408663
ICC via MPP: peak = 15.914254, avg = 2.527335, rms = 2.975906

In presence of noise, the -ql MPP now seems to be slightly better than the non-parametric ICC, with 50 training points.


And just for comparison, with 5000 noisy training patches:

ICC: peak = 7.416022, avg = 0.552852, rms = 0.679590
MPP -ql: avg err = 0.660334, max err = 4.727753
MPP -qm: avg err = 0.715241, max err = 5.930367
MPP -qh: avg err = 0.566079, max err = 5.764800


Regards,
Gerhard


Other related posts: