I'm curious how they're able to identify what parts of images are house numbers. To me, that seems like it would be a harder problem than determining what numbers are present. But I have really little knowledge in this domain so I might be completely wrong.