CNNs and MLPs trained on ~6k 15x15 optic flow patterns generated by uniformly sampled combinations of translation (T) and rotation (R).