Here is a potentially useful representation for face classification. While the article proposes a more general framework, it would be interesting to implement this in a domain-specific way. Any thoughts?