Why does colour.RGB_to_XYZ definition output differs from Bruce Lindbloom's calculator?

[blue-j]

I have a script that does a Mini Batch Kmeans clustering with sklearn and then calculates the XYZ value of each RGB cluster. Unfortunately, I cant find anywhere where the output can be sanctified. Bruce Lindbloom's calculator says I am wrong... and so I come here to see, have I made a mistake in the code?

    # Set variables to be used to convert to CIEXYZ
    wt = np.array(list(map(float, args.whitexy.split(','))))
    rd = np.array(list(map(float, args.redxy.split(','))))
    gr = np.array(list(map(float, args.greenxy.split(','))))
    bl = np.array(list(map(float, args.bluexy.split(','))))

    # this is D65; if D50 is desired, use 0.3457, 0.3585 for illuminant_XYZ
    illuminant_XYZ = np.array([0.3128, 0.3289])
    illuminant_RGB = wt
    chromatic_adaptation_transform = "Bradford"
    p = np.array([rd, gr, bl])
    matrix_RGB_to_XYZ = colour.normalised_primary_matrix(p, wt)

    # Cluster the input image
    clustered_image, labels, kmeans = cluster_image(args.input_image, args.num_clusters)
    total_pixels = labels.size
    res = []
    for i, color in enumerate(kmeans.cluster_centers_):
        rgb = np.array([color[0] / 255, color[1] / 255, color[2] / 255])
        hsv = colour.RGB_to_HSV(rgb)

        # if white points are equal, we do not use chromatic adaptation
        if np.allclose(wt, illuminant_XYZ):
            ciexyz = colour.RGB_to_XYZ(
                rgb,
                illuminant_RGB,
                illuminant_XYZ,
                matrix_RGB_to_XYZ,
            )
        else:
            ciexyz = colour.RGB_to_XYZ(
                rgb,
                illuminant_RGB,
                illuminant_XYZ,
                matrix_RGB_to_XYZ,
                chromatic_adaptation_transform,
            )

I am not getting any errors and the code looks good to me (it's much longer, but I have spared you!) but what i am concerned about is my execution of the RGB to XYZ function. sRGB is the source space.

this is what the args look like from the command line for an sRGB image:
"0.3128, 0.3289" "0.6488, 0.3304" "0.3211, 0.598" "0.1558, 0.0664"

for RGB (82, 115, 169) i get an XYZ of: 0.42641328 0.44800904 0.7002739

apparently, this is not correct? RGB (158, 180, 214) is the color reported from those XYZ coordinates (0.42641328 0.44800904 0.7002739) in Bruce's calculator.

I also tried of course to enter the RGB (82, 115, 169) with no adaptation, and Bruce's XYZ is: 0.167694 0.169187 0.399104

Thank you in advance for any time you may take to review this. : ) -J

[blue-j]

This is what the NPM looks like for this example:

[[ 0.36185485 0.39314477 0.19604934]
[ 0.1842738 0.73217244 0.08355376]
[ 0.01160077 0.09905142 0.97873668]]

[blue-j]

It seems I can calculate the CCTF if I know the primaries and white point in xy? this seems like the road to joy:

http://www.brucelindbloom.com/Eqn_RGB_XYZ_Matrix.html

Maybe you could incorporate these matrices that are missing from Colour? I may be playing the fool, forgive me, I am an amateur. - J

[KelSolaar]

Colour supports only a few RGB spaces in the CCTF encodings, notably missing Adobe RGB, Display P3, and many more. What route do you suggest? I am converting the input colors into 70+ spaces - all RGB spaces supported by Colour

A single colour component transfer function (CCT) can be used by many RGB colourspaces. But each RGB colourspaces has two, one for encoding one for decoding:

In [1]: import colour

In [2]: colour.models.RGB_COLOURSPACE_DISPLAY_P3
Out[2]:
RGB_Colourspace('Display P3',
                [[ 0.68 ,  0.32 ],
                 [ 0.265,  0.69 ],
                 [ 0.15 ,  0.06 ]],
                [ 0.3127,  0.329 ],
                'D65',
                [[  4.86570949e-01,   2.65667693e-01,   1.98217285e-01],
                 [  2.28974564e-01,   6.91738522e-01,   7.92869141e-02],
                 [ -3.97207552e-17,   4.51133819e-02,   1.04394437e+00]],
                [[ 2.49349691, -0.93138362, -0.40271078],
                 [-0.82948897,  1.76266406,  0.02362469],
                 [ 0.03584583, -0.07617239,  0.95688452]],
                eotf_inverse_sRGB,
                eotf_sRGB,
                False,
                False)

It seems I can calculate the CCTF if I know the primaries and white point in xy?

No, you can compute the normalised primary matrix (NPM) with them.

Maybe you could incorporate these matrices that are missing from Colour?

We have them all already!

[blue-j]

i hope i do not try your patience by asking what probably is a very basic question, but when do you apply the CCTF / EOTFs? how do you know when to use them? encoding, decoding... which?

[KelSolaar]

See Poynton's Gamma Faq, they are "Gamma Correction".

[blue-j]

Thank you for the resource! : )

-J