Unexpected behaviour with "colour.colorimetry.reshape_sd" definition

[frankkjensen]

I have an issue when reshaping some spectra. Below I load a spectrum with non-uniform wavelength intervals (file attached: White CRI98 6000K.csv)

When reshaping to (400, 700, 5) and (300, 700, 5) it works as expected, but when I try (300, 700, 5) the step is 5.033333, except for one point, where it is 2.36666.

import colour

sds = colour.read_sds_from_csv_file('White CRI98 6000K.csv', delimiter=',', header=None, names=['Wavelength', 'White CRI98 6000K'])
sd = sds['White CRI98 6000K']

wls = sd.wavelengths
wlSteps = [wls[i+1] - wls[i] for i in range(len(wls)-1)]
print(wls)
print('Min:', min(wlSteps), '\tMax:', max(wlSteps), '\tAverage:', sum(wlSteps)/len(wlSteps))
# The wavelength distribution is not uniform
print(sd.shape)
# Shape reports the minimum wavelength interval

sd_reshaped = colour.colorimetry.reshape_sd(sd, colour.SpectralShape(400, 700, 5))
wls = sd_reshaped.wavelengths
wlSteps = [wls[i+1] - wls[i] for i in range(len(wls)-1)]
print(wls)
print('Min:', min(wlSteps), '\tMax:', max(wlSteps), '\tAverage:', sum(wlSteps)/len(wlSteps))
print(sd_reshaped.shape)
# Works as expected

sd_reshaped = colour.colorimetry.reshape_sd(sd, colour.SpectralShape(300, 700, 1))
wls = sd_reshaped.wavelengths
wlSteps = [wls[i+1] - wls[i] for i in range(len(wls)-1)]
print(wls)
print('Min:', min(wlSteps), '\tMax:', max(wlSteps), '\tAverage:', sum(wlSteps)/len(wlSteps))
print(sd_reshaped.shape)
# Works as expected

sd_reshaped = colour.colorimetry.reshape_sd(sd, colour.SpectralShape(300, 700, 5))
wls = sd_reshaped.wavelengths
wlSteps = [wls[i+1] - wls[i] for i in range(len(wls)-1)]
print(wls)
print('Min:', min(wlSteps), '\tMax:', max(wlSteps), '\tAverage:', sum(wlSteps)/len(wlSteps))
print(sd_reshaped.shape)
# The wavelength interval is 5.033333, except for one point.

Best regards,
Frank

[KelSolaar]

Hi @frankkjensen,

The code tries to preserve as much as possible of the original data, there are some explanations in this issue but let me give you some more details. All the magic happens in the colour.SpectralDistribution.interpolate method.

Given the requested shape is (300, 700, 5), the code finds the minimum and maximum wavelength in the distribution of your spectrum: (397.07, 779.67)

• There is no data before 397.07, thus the next usable point for interpolation is 398
• There is no data beyond 779.67, thus the previous usable point for interpolation is 779
• We can thus safely interpolate between 398 and 779 on integer wavelengths
• The requested shape is (300, 700, 5) thus the effective interpolation shape becomes (398, 700, 5)

Such a shape will generate fractional intervals!

It is a case where I would recommend doing a direct extraction of the values you need as follows:

sd[colour.SpectralShape(300, 700, 5).wavelengths]

Hope it makes sense!

Cheers,

Thomas

[frankkjensen]

Ok, I understand.
I was trying to convert several spectra with different wavelengths to the same shape to be able to convert them to a multi-spectral distribution, which requires the same shape (right?). But really 99.9% of the times, I would chose 1nm steps (or an integer faction of 1nm), and that would work just fine since the start and end are rounded to integers, right?

Best regards,
Frank