[css-color] D50 mismatch in Lab ↔ sRGB algorithms

[danburzo]

In Section 17. Sample code for color conversions of the CSS Color Module Level 4 spec, there's a discrepancy with the D50 white point values.

In the XYZ ↔ Lab conversion code, it's assumed to be [0.9642, 1.0000, 0.8249] (as specified by the ICC spec).

However, the D50 ↔ D65 chromatic adaptation uses Bradford matrices from this page, which have been computed for D50 being [0.96422, 1.00000, 0.82521].

I think the correct approach is to use the ICC-defined D50 throughout, no?

(Also, I'm not clear on why the ICC value for the D50 white point differs from the generally accepted one. e.g. Matlab).

[danburzo]

It may be useful to specify in the section about lab() and lch() colors the specific values for the D50 white point as defined in the Profile Connection Space: [0.9642, 1.0000, 0.8249].

This document contains some matrices that may be useful.

[svgeesus]

Thanks for reporting the discrepancy!

neither ICC nor Lindbloom are authoritative for D50, which is defined by the CIE. I will check up and update the spec (and sample code) with the correct value.

I will also check in Wyszecki & Stiles when I get home.

[danburzo]

Thanks!

So if I understand correctly, the intent of the spec is to:

1. assume CIELAB is relative to D50 (as defined by CIE, not ICC); and furthermore,
2. the method of computing sRGB ↔️ CIELAB (in particular, the choice of matrices) is non-normative.

In regards to 1, this sentence:

D50 is also the whitepoint used for the profile connection space in ICC color interconversion,

I think may benefit from clarifying the distinction between the ICC definition of D50 and the CIE definition of D50.

In regards to 2, I'm a bit concerned that I've seen many variations on the matrices used to get sRGB ↔️ XYZ_D65 ↔️ XYZ_D50, or direct sRGB ↔️ XYZ_D50 pairs, and I wonder whether some well-defined matrices in the spec won't lead to more consistent implementations, since the approximations and the floating-point errors have a compounding effect on the result.

P.S. This page another set of matrices, and this sentence which further confuses things:

and the values of D65 and D50 from CIE 15.2 (and the ICC specification for D50) D65 = [95.04; 100; 108.89]; D50 = [96.42; 100; 82.49];

Not having access to CIE 15.2, I'm having a hard time piecing things together...

[svgeesus]

I've seen many variations on the matrices used to get sRGB ↔️ XYZD65 ↔️ XYZD50, or direct sRGB ↔️ XYZD50 pairs

Firstly, you can't do any of those with just a matrix; the sRGB needs to be linearized first. But having done that step, the operations for lin_sRGB to XYZ_D65, and XYZ_D65 to LMS, and white adjustment, and LMS to XYZ_D50 are all 3x3 matrices so the overall matrix is obtained by matrix multiplication.

The slightly different value for the Z component of D50 in ICC is a mistake, not an intentional change, but one they are now unable to correct without requiring all deployed CMS to be updated.

I feel that it would be better to express the correct values in the spec, with enough significant digits to avoid roundoff error.

[danburzo]

Also, in the Lab to LCH code, I think Math.atan2() should be normalized to the interval [0, 360).

[svgeesus]

Yes, indeed it should. Good catch. Although H is also defined to wrap, if someone enters a hue outside that range.

[danburzo]

Ah yes, I had somehow misread that H is <number> not <hue>. Too much time in front of the screen :)

[danburzo]

Some updates from my D50 saga: As I wrote in this thread, I was not able to find appropriate matrices for the conversion between sRGB and CIELab that did not introduce non-negligible errors. One particular doozy is that grayscale RGB colors will produce a Chroma of up to 0.0035 in the LCH equivalent.

[danburzo]

One last thing is nagging me: is there a benefit to keeping the CIELab values relative to D50 rather than sRGB's D65? Since the standard D50 does not match the values from ICC / Profile Connection Space... Defining lab() as relative to D65 would make computations simpler.

[svgeesus]

Yes, the value is that everyone else uses D50. If you use a spectrometer to measure color it will give you a readout in Lab with a D50 whitepoint.

[danburzo]

Thank you for the clarification!