Camera characterization gives "symetrical results"?

I downloaded the code from :

https://www.mathworks.com/matlabcentral/fileexchange/40640-computational-colour-science-using-matlab-2e?s_tid=srchtitle

I modified the code for my purposes. Please find all the attached elements to run the script at your end. What I don't understand is how the final resulting image looks exactly like the starting image? I understand how the 3rd degree polynomial is calculated but there must be something I miss in applying the M3 matrix back to my original image?

Any help is appreciated.

[Roger Breton]

I am no mathematician but I reasoned, once the M3 matrix is calculated, and the values of the camera responses converted to XYZ, before the Lab / DeltaE color differences are computed, why not "peek" into the XYZ values? So that's what I did in the following, by copying the XYZ values into a Temporary variable, for later convert to sRGB to display on my monitor :

M3=txyz'/trgb3';

pxyz3 = (M3*trgb3')';

pxyz3(pxyz3<0)=0;

Temp = pxyz3;

Original = xyz2srgb_D50(Temp);

Original255 = Original*255;

Original_img = reshape(Original, 14, 10, 3);

resized_orginal = imresize(Original_img, [14*200, 10*200], 'nearest');

rotated_image = rot90(resized_orginal);

flipped_vertical = flipud(rotated_image);

imshow(flipped_vertical);

The result is "perfect" :

Still. I'm unlcear as two what is that I'm doing that does not give me these results when I process the image through the same M3 matric?

[Roger Breton]

Well, ... I'll be damned. I went back to redo the RGB "patch extraction". I started off the very same RGB image I used for "validation" (DSC_0331 Assemblé Redressé 5x3.png). I ran the image through the 'patchmask' m file. Had to massage the data around in Excel to get it into a 140x3 form. But then I executed the script same as before, exactly, using this data. And, lo and behold, it looks like I'm getting the expected result, this time. Here is a screen capture of the final (beautiful!) "matched" image :

The results are not "perfect" when I inspect row 5 'neutral patches' but it's mighty encouraging.

Now the question is "Can I replicate this experiment?", just to be sure before I brag to my Photoshop students...

Answer 1

[Roger Breton]

Looks like there was a discrepancy between the RGB data used to train the model and the RGB data used to test the model. It's so much tedious work, at this stage? I was trying to avoid recreating everything from scratch... I was confident my data was "right" and it's only after having paid umpteen visits to ChatGPT and Twitter Grok, and scraping the web for possible references, that I decided in extremis to bite the bullet and recreate the original training data set again, painstakingly... Hope my mistake helps someone in the future.