It's not very good. I had that idea around the same time, tried to use MealSnap, gave up and decided the problem was on the wrong side of the difficult/interest

It's not very good. I had that idea around the same time, tried to use MealSnap, gave up and decided the problem was on the wrong side of the difficult/interesting balance for me. After looking at the photos I took for ~a week, I felt that as a human I'd be wildly inaccurate with the calorie counts, so how on earth could a machine be better? (Though, I was eating in a company cafeteria. Combining photos with geolocation for popular restaurants could be a good start to a more reliable, if sparse, approach.)