I think that is the strength of the deep learning methods that have been thrown at the problem in the meantime. They learn to pick a reasonable choice out of th

I think that is the strength of the deep learning methods that have been thrown at the problem in the meantime. They learn to pick a reasonable choice out of the null space by looking at the scene. Only good if the assumptions encoded into the network actually hold though. Kind of applies to us deep human networks as well. That's why optical illusions are fun to see.