Let us know what you find out and we can improve things if it is too painful.
Yeah I don't have my Ohno implementation vectorized either (MATLAB). It's a pain. Regarding numpy, yes I'm aware. One thing that gave me concern in the CAM16 code was the calls to other functions for every computational step. Function calls are cheap but not free and I've found that these can add up in well vectorized code. But it's definitely something that needs testing and examination before I really finish that assumption. As they say, premature optimization...
The other thing I'm trying to figure out, which is just a quality of life issue, is why when I run the tests any figure window tests block and have to be manually interacted with (closed) on windows before the next test can run.
Also wondering if I can configure tests to run in parallel.
(these are not necessarily pycolor issues, just me trying to get used to this dev ecosystem I'm running)
Sadly, I have other work to do today so I don't have as much time to toy with color. But my intention is to adopt it as the basis of my work and move away from Matlab. Mathworks licensing scheme recently really disappointed me and I can't continue to build any of my tools on their ecosystem. Too much risk in the licensing for commercial use.
I have vectorised most of our Ohno (2013) implementation and made some optimisations, e.g. caching, etc...: colour-science/colour#951
colour.temperature.uv_to_CCT_Ohno2013definition is ~100x faster.colour.temperature.CCT_to_uv_Ohno2013definition is ~425x faster.
It is merged in
develop.
pandas will let you manipulate the data layout as you wish to output it the way you want
otherwise it's kind of a virtual spreadsheet
# 2. FAIRCHILD, Mark D. et PIRROTTA, Elizabeth.
# Predicting the lightness of chromatic object colors using CIELAB.
# Color Research & Application, 1991, vol. 16, no 6, p. 385-393.
# sample | Munsell notation | L* | C* | h* | Observed L*
raw_data_2 = """
1 "5R3/2" 30.58 13.68 25.46 38.1
2 "5R3/6" 31.19 33.33 25.34 39.1
3 "5R3/10" 30.89 53.40 24.60 38.8
4 "5R5/2" 51.64 9.95 26.19 55.2
5 "5R5/8" 51.41 40.99 27.15 56.7
6 "5R5/14" 52.00 71.23 29.40 62.9
7 "5R7/2" 72.49 8.72 28.37 71.4
8 "5R7/6" 71.56 29.51 28.29 72.1
9 "5R7/10" 72.60 48.64 27.81 75.0
10 "5Y3/1" 30.34 8.30 90.99 36.3
11 "5Y3/2" 30.30 13.97 89.36 36.4
12 "5Y3/4" 30.58 28.50 87.52 35.4
13 "5Y5/2" 51.52 15.87 93.54 57.0
14 "5Y5/6" 51.43 47.03 89.96 48.6
15 "5Y5/8" 51.59 59.71 88.42 50.9
16 "5Y7/2" 71.77 16.67 95.07 71.4
17 "5Y7/8" 71.82 59.83 91.06 65.5
18 "5Y7/12" 71.69 86.70 89.59 66.1
19 "2.5G3/2" 30.22 10.95 151.02 36.4
20 "2.5G3/6" 30.81 34.31 155.69 39.2
21 "2.5G3/10" 30.91 55.44 156.91 40.9
22 "2.5G5/2" 51.20 12.60 149.41 56.0
23 "2.5G5/8" 52.11 48.61 152.67 59.5
24 "2.5G5/12" 50.81 65.89 154.59 61.5
25 "2.5G7/2" 72.79 13.55 148.52 71.3
26 "2.5G7/6" 72.16 37.75 150.32 73.4
27 "2.5G7/10" 71.32 58.28 152.88 70.0
28 "5PB3/2" 30.27 9.99 275.89 38.4
29 "5PB3/6" 30.41 27.82 274.26 44.4
30 "5PB3/10" 30.42 44.05 276.56 48.6
31 "5PB5/2" 51.82 7.20 273.99 57.1
32 "5PB5/8" 52.05 31.87 272.63 61.4
33 "5PB5/12" 51.14 46.26 270.92 62.8
34 "5PB7/2" 72.08 6.05 266.63 75.4
35 "5PB7/6" 71.87 22.02 267.04 73.3
36 "5PB7/8" 71.17 28.68 266.47 74.7
"""
BL_data_2 = pd.read_csv(StringIO(raw_data_2), sep=" ", index_col=0, header=None, names=["sample", "Munsell", "L", "C", "h", "Ln"])
# Convert to xyY to match the other dataset
Lab = colour.LCHab_to_Lab(np.vstack([BL_data_2["L"].to_numpy(), BL_data_2["C"].to_numpy(), BL_data_2["h"].to_numpy()]).T)
XYZ = colour.Lab_to_XYZ(Lab)
xyY = colour.XYZ_to_xyY(XYZ)
BL_data_2["Y"] = xyY[:, 2] * 100
BL_data_2["x"] = xyY[:, 0]
BL_data_2["y"] = xyY[:, 1]
Labn = colour.LCHab_to_Lab(np.vstack([BL_data_2["Ln"].to_numpy(), BL_data_2["C"].to_numpy(), np.zeros_like(BL_data_2["h"].to_numpy())]).T)
XYZn = colour.Lab_to_XYZ(Labn)
BL_data_2["Yn"] = XYZn[:, 1] * 100
# The weight is the number of observers in the study : n = 11
BL_data_2["w"] = 11
BL_data_2# 3. Withouck, Martijn & Smet, Kevin & Ryckaert, W. & Pointer,
# M. & Deconinck, Geert & Koenderink, Jan & Hanselaer, Peter. (2013).
# Brightness perception of unrelated self-luminous colors.
# Journal of the Optical Society of America.
# A, Optics, image science, and vision. 30. 1248-55.
# 10.1364/JOSAA.30.001248.
# 9 observers, reference radiometric luminance = 51 Cd/m²
# Results given in CIE LUV 1976 10° observer
# Columns : huv;10 | suv;10 | Qgeom
raw_data_3 = """
291.82 0.53 54.65
289.04 0.86 59.82
287.04 1.23 62.50
286.63 1.71 69.19
62.70 0.21 48.71
69.75 0.40 46.27
78.04 0.65 48.81
79.02 0.94 58.39
77.21 1.12 58.46
17.62 0.21 50.57
346.85 0.57 55.12
343.30 0.89 55.54
341.18 1.26 68.07
339.18 1.83 73.26
83.22 0.37 49.30
92.23 0.58 44.79
97.46 0.79 50.24
97.34 0.95 52.34
96.73 1.14 61.49
223.45 0.63 52.63
223.89 0.78 57.00
224.03 1.02 58.24
224.11 1.34 62.56
225.92 1.66 65.87
36.05 0.46 48.19
31.29 0.88 51.98
29.33 1.23 46.60
29.24 1.55 52.18
28.24 1.98 64.98
153.58 0.55 50.58
155.98 0.62 50.73
159.48 0.92 55.17
161.38 1.18 60.36
162.01 1.44 61.88
101.66 0.19 47.57
139.55 0.75 47.70
143.33 1.14 60.10
145.02 1.54 68.18
139.03 1.97 76.04
44.00 0.16 49.72
15.68 0.78 50.88
13.12 1.30 58.11
11.99 1.99 66.35
11.40 2.74 64.15
12.06 3.61 79.12
264.89 0.45 51.64
258.16 1.43 61.18
257.34 2.05 61.69
256.10 2.69 71.59
252.28 3.04 76.95
195.84 0.29 51.56
201.09 0.60 50.85
202.26 0.84 60.38
203.30 1.30 66.12
49.52 0.13 48.28
49.45 0.54 47.58
97.94 0.00 50.00
41.79 1.02 48.49
"""
BL_data_3 = pd.read_csv(StringIO(raw_data_3), sep=" ", index_col=None, header=None, names=["h", "s", "Yn"])
BL_data_3["Y"] = 0.51 * 100
BL_data_3["H"] = BL_data_3["h"] / 180 * np.pi
# Convert s, h to UV in CIE LUV 10°
BL_data_3["U"] = -1 / 13 * BL_data_3["s"] * np.abs(np.cos(BL_data_3["H"]))
BL_data_3["V"] = -np.abs(1 / 13 * BL_data_3["s"] * np.abs(np.cos(BL_data_3["H"])) * np.tan(BL_data_3["H"]))
# Because tan(h) = V / U == -V / -U, the tan function will always yield results in ]-pi / 2; pi / 2[
# if hue is in [180; 270]°, U and V are expected both negative and we need to rotate the result by 180°
# notice that .where() replaces values that have False in the corresponding row, so the logic is reversed
U_positive = BL_data_3['h'].between(90, 270, inclusive="both")
BL_data_3["U"].where(U_positive, -BL_data_3["U"], inplace=True)
V_positive = BL_data_3['h'].between(180, 360, inclusive="both")
BL_data_3["V"].where(V_positive, -BL_data_3["V"], inplace=True)
# Get u'v' from UV
D65_10 = colour.colorimetry.CCS_ILLUMINANTS['CIE 1964 10 Degree Standard Observer']['D65']
D65_10_uv = colour.xy_to_Luv_uv(D65_10)
BL_data_3["u"] = BL_data_3["U"] + D65_10_uv[0]
BL_data_3["v"] = BL_data_3["V"] + D65_10_uv[1]
# Convert to xy
uv = np.vstack([BL_data_3["u"], BL_data_3["v"]]).T
xy = colour.Luv_uv_to_xy(uv)
BL_data_3["x"] = xy[:,0]
BL_data_3["y"] = xy[:,1]
# The weight is the number of observers in the study : n = 9
BL_data_3["w"] = 9
BL_data_3
@KelSolaar: is there a way to convert CIE XYZ 1964 10° to CIE XYZ 1931 2° ?
because the last one and the next one use CIE 1964 10°
This definition can define the outer surface for example: https://github.com/colour-science/colour/blob/develop/colour/volume/spectrum.py#L384
Then you would need to produce spectra inside that volume, and finally generate the transform between the two datasets.
I had started some work around that a while ago, might be some stuff to dig in that Colab Notebook: https://colab.research.google.com/drive/1YO6kfohVxjdGm4t6I3JMifff00BB2SuM?usp=sharing
4 replies
There is also the paper from Richard Kirk from Filmlight that might be relevant here where they map to CIE 2006.
4 replies
Kirk got fair-enough results with that
hmm, now I wonder how Hellwig, Fairchild and Nayatani can compare the B/L datasets if some of them are defined with chromaticities in 2° 1931 and some other in 10° 1964
so far, I have just used xy regardless of the declaration space
These are the 3 datasets above, x is predicted brightness with HKE, y is experimental data
Oh sorry, I just saw your message
@KelSolaar: I will update my Collab with the fittings and datasets when I'm done, you will be able to grab the data from there as you want
that's a total of something like 125 observers and 180 data points
@KelSolaar: I don't know if you followed, but MacAdam moments actually make a great job at predicting HKE by treating brightness as the invert of saturation
I haven't looked too much yet but I saw some interesting stuff in MacAdam (1938), e.g. https://imgur.com/iqbHme4
def normalise_multi_signal(multi_signal):
multi_signal = multi_signal.copy()
f_n = colour.sd_to_XYZ(colour.sd_ones(), cmfs=multi_signal) / 100
print('"{0}": {1}'.format(multi_signal.name, f_n))
multi_signal.values = multi_signal.values / f_n
return multi_signal
CMFS_1_NAME = 'CIE 1931 2 Degree Standard Observer'
CMFS_1 = normalise_multi_signal(colour.colorimetry.MSDS_CMFS[CMFS_1_NAME].copy().align(SHAPE))
# Generate a LUT of test XYZ vectors
XYZ_1931 = []
for X in np.arange(0, 1, 0.05):
for Y in np.arange(0, 1, 0.05):
for Z in np.arange(0, 1, 0.05):
XYZ_1931.append(np.array([X, Y, Z]))
# Convert XYZ 1931 vectors to spectra
SD = []
for XYZ in XYZ_1931:
SD.append(colour.XYZ_to_sd(
XYZ, cmfs=CMFS_1, method='Jakob 2019', optimisation_parameters={'options': {
'ftol': 1e-5
}}))
# Convert back spectra to XYZ 1931 2°
XYZ_1931 = colour.msds_to_XYZ(np.array(SD), method='Integration', cmfs=CMFS_1, shape=SHAPE) / 100AttributeError: "SpectralDistribution.__iter__" object has been removed from the API.
so the RMSE between 1964 and 1931 observers for arbitrary spectra is 0.00050
2 replies
converting XYZ from 1964 to 1931 with a matrix
array([[ 1.06756581, 0.07134588, -0.13992265],
[-0.06910109, 0.988538 , 0.13924861],
[ 0.00140968, -0.05954396, 0.99820504]])drops it to 0.00016
with the XYZ grid, I upsampled the spectra with Jakob & Hannika 2019, then converted spectra to XYZ 1931 and 1964