What do you do when you’ve written a funky new shape recognition engine? Throw it at an impossible problem, of course.

I was vehement that we’d never be able to recognise text in real-world photographic scenes. (These days this is commonplace, but back in 2001 it was still a difficult research problem.) I was extremely surprised when I proved myself wrong.

Where Foveola was satisfyingly elegant, SceneReader was messy, heuristic and approximate – just like reality. I think the only reason it worked so well (for its time) was that I knew bugger-all about machine vision and image processing when I started – so I didn’t get sunk by the existing preconceptions.