This is the supplementary materials for the submission number 3765 ("Learning Representation in Colour Conversion") to ICLR-2020.



The content of this folder: 

 - "transformation_matrices"
   This directory contains the linear transofrmations we computed for each lesion.
   We conducted this study only for three networks with K=8 and D=128 whose input is in RGB colour space and their outputs either RGB, DKL or LAB.
   In instructions below:
   <XXX> can be one of these three output colour spaces (dkl, lab or rgb).
   <Y> is the vector index ranging from 0 to 7 corresponding to each embedding vector.

   
   * "<XXX>DeltaSelectionThreshold.txt"
     Defines the threshold to choose the most changed pixels.

   * "<XXX>L<Y>AllPixels.txt"
     The transformation matrix computed based on all pixels of an image.
     (3x3 matrix)

   * "<XXX>L<Y>MostChangedPixels.txt"
     The transformation matrix computed based on 1% of pixels.
     (3x3 matrix)
     
   * "<XXX>RSQ_AllPixels.txt"
     R-squared as a measure of goodness of fit.
     There are 8 rows, one for each lesion. The two columns indicate (1) the rsquared computed on the correlation between the pixels Lab values (L,a,b, together in a single vector) of the no-lesion images and the same pixels of the lesioned images (2) between the pixels of the lesioned images and the ones of the transformed no-lesion images.

   * "<XXX>RSQ_MostChangedPixels"
     Similar as above for most changed pixels.
