Will continue the discussion there!
@Saransh-cpp : Yes will do!
Hi, I am Dani, nice to meet you all. I am just beginning to try colour-science for the first time. Particularly I'm interested in colour.colour_correction() function, and wanted to ask for some help. Concluding from the documentation, does those function take samples of colours and trying to interpolating the colours to create 3D LUT?
As the proof of concept, I got a normal photo of ColourChecker and another same image that I edited to make it more 'cool' (as in colour temp), and extracted colors from each square of both images. Then I'm trying to generate LUT for my low temp image back to the original using color_correction(). I got the results were 'close' to the original, but not exact. Does this is expected (from comparing image to image), rather than comparing to the actual RGB color of the chart?
If anybody want to check, this is my source : https://gist.github.com/otivedani/81d50b0ab3a959a02e7b60d2f6f43941 (colab link available). Thank you in advance
If anybody want to check, this is my source : https://gist.github.com/otivedani/81d50b0ab3a959a02e7b60d2f6f43941 (colab link available). Thank you in advance
Hi Dani!
Concluding from the documentation, does those function take samples of colours and trying to interpolating the colours to create 3D LUT?
Not quite, depending on the variant they either do a linear transformation via a 3x3 matrix that wil rotate and scale the colours to the best fit, or do something similar but using a 3xn polynomial to distort the space to the best fit
Thus, there is no data discretization/3D LUT involved, only functions.
The functions will not be able to fully fit the target and source spaces so it is expected to have some differences, worth noting that polynomials methods might be subject to "explosions" outside the domain they are used on. Put another way, they are very good at interpolating the colours defined by the target and source spaces but extrapolation outside that is bound to behave very unexpectedly.
Hi @KelSolaar , thank you for helping me to understand! I will get to learn more from the thread too. In other words, the functions take 3 -> 3 (as in 3D, rather than 1D each channel), right?
What I mean to create 3D LUT is, I have an idea to apply the correction to a neutral 3D LUT image rather than the source image itself, like this : https://streamshark.io/obs-guide/converting-cube-3dl-lut-to-image , so the adjustment could be reused. Did you think it is going to work?
What I mean to create 3D LUT is, I have an idea to apply the correction to a neutral 3D LUT image rather than the source image itself, like this : https://streamshark.io/obs-guide/converting-cube-3dl-lut-to-image , so the adjustment could be reused. Did you think it is going to work?
Hey @KelSolaar,
I wanted to work on #796 and just required a little guidance on what information and output format needed.
I wanted to work on #796 and just required a little guidance on what information and output format needed.
Hi, I need some help regarding using "colour_correction".
I have a x-rite color chart with 24 colors, and I have found these values online as the reference:
reference_colors = [[115,82,68],[194,150,130],[98,122,157],[87,108,67],[133,128,177],[103,189,170],
[214,126,44],[80,91,166],[193,90,99],[94,60,108],[157,188,64],[224,163,46],
[56,61,150],[70,148,73],[175,54,60],[231,199,31],[187,86,149],[8,133,161],
[243,243,242],[200,200,200],[160,160,160],[121,122,121],[85,85,85],[52,52,52]]
and I have read the values from a raw image from my custom camera:
extracted_colors = [[60,54,59],[131,93,86],[50,60,83],[40,44,41],[82,76,108],[88,134,147],
[158,95,68],[40,49,90],[128,57,54],[41,31,46],[99,117,71],[172,120,72],
[35,43,80],[45,71,58],[113,43,39],[188,147,72],[136,65,84],[44,80,118],
[213,211,212],[146,146,150],[96,96,100],[58,58,63],[38,38,45],[28,30,39]]
and I want to use it like this:
corr_color = colour.colour_correction(extracted_colors[k], extracted_colors, reference_colors, method='Finlayson 2015')
I am however in doubt about the need to normalize them from [0,255] to [0,1] range and about the need to linearize
I have a x-rite color chart with 24 colors, and I have found these values online as the reference:
reference_colors = [[115,82,68],[194,150,130],[98,122,157],[87,108,67],[133,128,177],[103,189,170],
[214,126,44],[80,91,166],[193,90,99],[94,60,108],[157,188,64],[224,163,46],
[56,61,150],[70,148,73],[175,54,60],[231,199,31],[187,86,149],[8,133,161],
[243,243,242],[200,200,200],[160,160,160],[121,122,121],[85,85,85],[52,52,52]]
and I have read the values from a raw image from my custom camera:
extracted_colors = [[60,54,59],[131,93,86],[50,60,83],[40,44,41],[82,76,108],[88,134,147],
[158,95,68],[40,49,90],[128,57,54],[41,31,46],[99,117,71],[172,120,72],
[35,43,80],[45,71,58],[113,43,39],[188,147,72],[136,65,84],[44,80,118],
[213,211,212],[146,146,150],[96,96,100],[58,58,63],[38,38,45],[28,30,39]]
and I want to use it like this:
corr_color = colour.colour_correction(extracted_colors[k], extracted_colors, reference_colors, method='Finlayson 2015')
I am however in doubt about the need to normalize them from [0,255] to [0,1] range and about the need to linearize
from what I understand the reference values I found online should be sRGB, so I expect I need to linearize them?
and is that with the "colour.models.eotf_inverse_sRGB()" or "colour.models.eotf_sRGB()"?
and I believe my image is pure raw and doesnt have any gamma encoding
and is that with the "colour.models.eotf_inverse_sRGB()" or "colour.models.eotf_sRGB()"?
and I believe my image is pure raw and doesnt have any gamma encoding
So it looks like you are using sRGB 8-bit values here, so you must convert the array to float, divide by 255 and then apply the
eotf_sRGB to decode
>>> colour.models.eotf_sRGB(colour.utilities.as_float_array([122, 122, 122]) / 255)
array([ 0.19461783, 0.19461783, 0.19461783])thanks,
yeah, I also got to that conclusion after seeing examples and comments from different places.
I ended up just using the reference colors from 'colour' since I could see my x-rite chart was from January 2014:
D65 = colour.CCS_ILLUMINANTS['CIE 1931 2 Degree Standard Observer']['D65']
REFERENCE_COLOUR_CHECKER = colour.CCS_COLOURCHECKERS['ColorChecker24 - Before November 2014']
REFERENCE_SWATCHES = colour.XYZ_to_RGB(
colour.xyY_to_XYZ(list(REFERENCE_COLOUR_CHECKER.data.values())),
REFERENCE_COLOUR_CHECKER.illuminant, D65,
colour.RGB_COLOURSPACES['sRGB'].matrix_XYZ_to_RGB)Based on the example from https://github.com/colour-science/colour-checker-detection/blob/develop/colour_checker_detection/examples/examples_detection.ipynb
I also tried converting the reference I previously had from X-rite by doing what you said (I had also tried the cctf_decoding()/encoding() as in the example above) and I could see that the values where quite close but not really an exact match...
Is that because of all the colorspace conversions (xyY->XYZ->RGB)?
I can also see that the CCM really improves my images, even though it is not perfect
The images are white balanced in advance and the gain and exposure time has been adjusted in order to maximize the dynamic range while avoiding clipping
Are there any more steps I should apply before the correction which can improve the results further?
I can see that the biggest improvement to my images by applying the CCM is in the color saturation
The images are white balanced in advance and the gain and exposure time has been adjusted in order to maximize the dynamic range while avoiding clipping
Are there any more steps I should apply before the correction which can improve the results further?
I can see that the biggest improvement to my images by applying the CCM is in the color saturation
You could use Finlayson (2015) but it is really highly conditioned by the input data and here a ColorChecker 24 is often not enough
Here is a relevant post we wrote few weeks ago: https://www.colour-science.org/posts/the-colorchecker-considered-mostly-harmless/
quicker might be to cull the spectral data and save it as two different CSV files and load them with colour directly
@KelSolaar I am a little confused about where the code for this would go. I see we currently have a function
read_spectral_data_from_csv_file. Are there any other types other than sekonic and UPRTek that were already implemented that I can have a look at? If not, where should I fit the new parser?
We would probably add two new parsers, one for Sekonic and one for UPRTek, the latter sharing most code from the former as they are for practical purposes the same.
Can we move this discussion to the issue though, it will be easier to track down later?
@KelSolaar we have a custom camera (machine vision camera) which is being used for live sports streaming (ice hockey) and the goal is basically to make the picture look good, more true vibrant colors (we have actually several cameras, so they also have to look similar).
We can definitely see that we are missing on the color saturation and that is mostly what the CCM I tried generating seems to correct.
Because of the live requirement going to CIE Lab or something is a bit too costly, so that is why it would be preferable to keep in the native RGB or the YUV we transform to before encoding to H264.
I had the impression that when using degree=1 all methods where pretty much equivalent? And that Finlayson differed when going to higher orders?
And because it is indoors sports the illumination won't change, so I expected that doing the calibration once (like I already do for white balance and exposure) would work...
We can definitely see that we are missing on the color saturation and that is mostly what the CCM I tried generating seems to correct.
Because of the live requirement going to CIE Lab or something is a bit too costly, so that is why it would be preferable to keep in the native RGB or the YUV we transform to before encoding to H264.
I had the impression that when using degree=1 all methods where pretty much equivalent? And that Finlayson differed when going to higher orders?
And because it is indoors sports the illumination won't change, so I expected that doing the calibration once (like I already do for white balance and exposure) would work...
it is indoors sports the illumination won't change, so I expected that doing the calibration once (like I already do for white balance and exposure) would work...
Yes! This is almost the ideal scenario, ideally you would sample the illumination with a spectrometer, e.g. Sekonic C7000, measure the chart reflectances with another one, e.g. X-Rite I1 Pro, from there you could generate the reference chart values under the illumination of the location and calibrate the camera against that.
@KelSolaar so if I want the camera to be able to reproduce what an expectator in the stadium is seeing I would have to measure the reference chart values under that illumination with a X-Rite I1 Pro and preferably do the calibration and correction under a perceptual uniform space like CIE Lab.
But I don't get why I also need to "sample the illumination with a spectrometer, e.g. Sekonic C7000"? (Does the "X-Rite I1 Pro" requires those values?)
And, if like now, I use the reference values from D65, I would be "distorting" the picture to look like it was taken under such illumination? That might not be a very big problem for me if the picture is pretty, even though it might not be exactly the same color a person would see live. Or is there more I should be aware of?
I have read the blog post you mentioned, but as my camera is a custom one it hasn't been "carefully color calibrated", this is what I am trying to achieve in some sense ;)
But I don't get why I also need to "sample the illumination with a spectrometer, e.g. Sekonic C7000"? (Does the "X-Rite I1 Pro" requires those values?)
And, if like now, I use the reference values from D65, I would be "distorting" the picture to look like it was taken under such illumination? That might not be a very big problem for me if the picture is pretty, even though it might not be exactly the same color a person would see live. Or is there more I should be aware of?
I have read the blog post you mentioned, but as my camera is a custom one it hasn't been "carefully color calibrated", this is what I am trying to achieve in some sense ;)
In that case it would be to generate reference chart values with the illumination that you have measured in the indoor hockey location.
then with the chart captured by the cameras under that illumination you can correct much more precisely because the values are for the actual location, not an unrelated illuminant.
You can measure ambiant light with an i1 Pro but it might not be the best too for the job, it might work though, there is a head for that: http://cdn.northlight-images.co.uk/wp-content/uploads/2018/12/ambient-head.jpg