There is a lot of academic work in this area that's relevant, dating back at least 10 years. I'm surprised it wasn't mentioned.

Maximizing the entropy of projected normals works well, for example, like this paper by Polonsky et al.: http://www.cs.technion.ac.il/~gotsman/AmendedPubl/Oleg/Polon...

A related problem is automatically figuring out the "natural orientation" of a mesh, so your automatic renderings don't look upside down by accident. Fu et al. use pretty simple machine learning tools: http://www.cs.ubc.ca/nest/imager/tr/2008/upright_orientation...