Hacker News Reader

Comment by nadam | Hacker News Reader

This is a bit impractical. Recognizing the food with few errors would be horribly hard. I think it is sometinhg like speech recognition or automatic translation: yes, it is easy to create some simple not very good speech recognizer or translator software, but it is incredibly hard to create a high quality one. The only remaining option is to create a user interface where the user can type-in his daily food intake very easily/intuitively. Maybe most users would want to do this on their computer's screen and not on a phone. Of course there are plenty of calory-tracker, 'diet organizer' software on the market like http://www.dietorganizer.com/. But I am not an expert in this market.

View on HN

> Recognizing the food with few errors would be horribly hard.

I wonder if it would be possible to have some sort of a hybrid system, where automatic recognition is attempted, with a fall-back to manual input where this fails. (The manual input could be to select from several automatic "guesses") If the app stored all the data from all users, this could eventually lead to a really big database, and the automatic recognition might get better and better.

EDIT: Also, automatic vegetable/fruit recognition is actually in production in some supermarkets, for weighing/pricing. (I saw this in France.)

How will you know when you should fall back to manual? There's really no measurement of accuracy that I can think of. If you've reached the point in the algorithm where you know you've incorrectly identified the food it seems like you've already solved the problem.

That is a pretty common problem in machine learning. There are tons of solutions, but they are usually classifier specific. For example, if using a decision tree, one can take test data and determine for each leaf node in the tree what fraction of instances are classified correctly. For leaf nodes with accuracy below X, change the output of that node to "give up".

> EDIT: Also, automatic vegetable/fruit recognition is actually in production in some supermarkets, for weighing/pricing. (I saw this in France.)

I saw this in France. It sucked. I wish I'd documented it better, but as far as I could tell, the algorithm was "display the top five fruit/veg with a similar hue," i.e. you put on a courgette(zuchini) and it offered you--cucumber, lettuce, green pepper, cabbage, grapes. Pinnacle of machine vision algorithms, it ain't.

I disagree people want to do this on a computer screen. Pulling out an iphone is far less friction than remembering to enter this information next time you're on a computer.

Reply on news.ycombinator.com