PR: Implement support for Y'CbCr matrices generation.

[KelSolaar]

References #486.

[coveralls]

Coverage increased (+0.0003%) to 99.721% when pulling 2181c93 on feature/ycbcr_matrices into d716e14 on develop.

[KelSolaar]

@nick-shaw : Might be worth one of your eyes! :)

[nick-shaw]

Couple of points and/or references: in Supplement 18 of ITU-T (page 24) a matrix with 14 digits after decimal point is given and is to be used for Y′CbCr to R′G′B′.

This function would be able to calculate those 14 decimal place coefficients (or higher precision if desired). The 14 decimal place coefficients are just the 4 decimal place constants from BT.2020 run through the weighting to matrix calculations.

[nick-shaw]

>>> np.set_printoptions(suppress=True, formatter={'float_kind':'{:0.14f}'.format})
>>> matrix_YCbCr(K=WEIGHTS_YCBCR['ITU-R BT.2020'])
array([[1.00000000000000, -0.00000000000000, 1.47460000000000],
       [1.00000000000000, -0.16455312684366, -0.57135312684366],
       [1.00000000000000, 1.88140000000000, 0.00000000000000]])

This is exactly the matrix shown in the document.

[nick-shaw]

Looks good to me, except perhaps the docstring should be more explicit that the returned matrix is the Y'CbCr to R'G'B' matrix, not the R'G'B' to Y'CbCr matrix. Other than that I am happy to approve the PR.

[KevinJW]

I haven't fully looked at this, but I wonder if ITU T-REC-H.273 might have something to say on these, at least it states the expected behaviours for a number of the combinations. It does appear to be attempting to be definitive: This Specification can be used to provide universal descriptions to assist interpretation of video and image signals following decoding or to describe the properties of these signals before they are encoded

Kevin

[nick-shaw]

Not sure I completely follow, @KevinJW. Are you suggesting we adopt the "Video Code Points" values as an (optional?) way of passing the K and is_legal arguments to the functions? This feels like it is more a part of the metadata layer that we do not yet have. Ideally the functions would also return metadata describing how the result was encoded.

[nick-shaw]

BTW, can you add Kodak Photocd quantization?

The RGB_to_YCbCr function can already use arbitrary output scaling and offsets, using the out_range argument. See here for details. So the PhotoCD quantisation could be achieved using that together with the out_int argument.

[KelSolaar]

Thanks for the review @nick-shaw! I addressed the docstring notes and rebased on-top of develop. I will merge after the tests pass.

[KevinJW]

The numbers in that line in FFmpeg would be my "fault", can't remember exactly what source I used for it

Kevin