Learning to See From 3D Data