Maybe I wasn't clear. The spectral locus for each camera was derived from spectral stimuli 5 nm apart. The gamut boundary was not derived for one set of color patches. That was the training set. Looking at the manufacturers compromise matrices, the ones derived are not out of the question. I'm working on other training sets, but I'm not expecting to find compromise matrices that provide accurate color for those sets and completely fill the horseshoe, in the case of the cameras used for the above tests.
Jim
Jim,
I don't follow. If not from the Colorchecker patches from what did you derive the conversion matrix? I assumed they were derived to minimize some function of dE and then the gamut boundaries are just applying the matrix to the camera curves.
My point was that the Colorchecker, apart from having only 19 patches (the neutrals are effectively redundant), is a limited universe of colors. And even amongst that universe, there are large spectra variations that produce the exact same CIELAB and hence xy chromaticities.
I posit that if one examines other possible spectra for each of the colorchecker patches, without changing their Lab color, that it would be pretty easy to find matrixes that produced a gamut with portions overlapping beyond any arbitrary section of the horseshoe.
Basically, I don't believe the gamut of a camera sensor assy. can be defined by a xy outline of any sort unless the colors and illuminant, from which some sort of "optimal" matrix is generated, are constrained. Of course once any given matrix is generated, the response gamut, using that matrix, is well defined which is what your post shows.