What is a good composition? There are many theories, to take one, consider Arnheim's book "The Power of the Center" which proposes that radial organization competes with Cartesian organization to produce low energy forms that look stable and pleasing.
Where do these two types of organization arise? One answer is that they come from common patterns of optic flow as discussed by Gibson, e.g. see the pictures below.
Above we see a radial optic flow as seen by a pilot landing a plane. The same thing happens as we move forward anywhere really. The radial movement of things happens very naturally as we move forwards, so when we see things move radially in a still image, maybe we imagine we are moving forwards. An image which makes us imagine we are doing things is probably a more interesting one for us. What about the Cartesian grid?
My theory (someone may have said this before, I need to look it up) is based on the following thought experiment, although I want to wear a GoPro on my head for a week and collect this data, day and night. Take all the images we see over normal ecological life, such as this one when I'm lying on the sofa having a rest.
Then blur these, segment them, and consider the main forms based on recognition. We see for a start that things are skewed, its mainly diagonals. Also, in this case there is a big form that takes up a lot of the view (the cushion) and some lines pointing to the center (fireplace and right of cushion) etc, etc... Imagine an algorithm that takes all the frames obtained over a week and produces images which contain segments, ignoring high frequency components (although high frequency components could be encoded as some overall segment specific value if you wish). Then cluster these images using e.g. t-SNE. I predict that what would come out would be a set of archetypical major compositional forms that have over the centuries been discovered by artists, and which provide us with clear unconscious affordances. Compositions will differ based on whether one is looking at things close up or far away, inside or outside, lying down or walking, etc. Would existing categories of composition like the rule of thirds etc... come out ecologically (an Arnheimesque interpretation) or would they be entirely conventional and arbitrary (a Goodmanesque interpretation)? I predict the former would be the case. Anyone want to help me with this experiment? After doing this experiment, we could train a neural network to generate compositional frameworks from the experienced data, and then Arnheim 2.0 might get some reward for arranging its paintings according to one of these natural compositional formats.