
The mondrian.jpg image yields the best results with the following
constructor parameters:

WITH IMAGE NORMALIZATION OFF:

rpg = RegionProposalGenerator(

               sigma = 0.8,
               max_iterations = 40,
               kay = 5,
#               image_normalization_required = True,
               image_size_reduction_factor = 4,
               min_size_for_graph_based_blobs = 4,

               color_homogeneity_thresh = [20,20,20],      
               gray_var_thresh = 50,            
               texture_homogeneity_thresh = 120,           
               max_num_blobs_expected = 8
      )




A special challenge with the mondrian.jpg image:

The mondrian image in this directory did not work for out for
illustrating the RegionProposalGenerator module because the colors
used by Mondrian are all very saturated.

So when I convert the image into black and white, the gray level 
associated with every pixel is close to 255, which screws up 
the pairwise comparisons between the different regions in the
image.

THE ABOVE statement was based on using the generic "RBG" to "L"
converter for the Image class.  

HOWEVER, if I use the PyTorch's converter I get fairly acceptable
results as shown in this directory.


