
We tried multiple approaches for deriving convolutional masks from objects models. Maximally informative and maximally correlated masks require us to solve ill-posed optimization problems. Small windows sampled from the object model are not informative.