← Back to context

Comment by Nk26

4 days ago

Does the photo recognition attempt to carb count what it sees? Is that even possible? My son is a T1D and he still struggles with carb counting.

15 comments

Nk26

Reply
jrflowers  4 days ago

It is absolutely not possible to carb-count through photo recognition in a way that is reliable enough for a diabetic to safely use to make treatment decisions.

  • nihzm  4 days ago

    Indeed, around 2019 I was reading many computer vision papers for volume estimation and came across a few that tried to estimate the weight of the meals from pictures using the size of known objects (cutlery beside the plate). The idea was good but they were very far from accurate and not robust at all, and that was just for the weight, not even carb counting. I know CV is a fast moving field but I wouldn't bet that the tech has improved enough to be anywhere near medically safe.

  • Nk26  4 days ago

    I meant more in a general way, like a piece of pizza is usually around X carbs. We have apps that make the guess a bit easier but it's almost always a guess. I was thinking could this look at a photo and know there's a sweet potato, a piece of chicken and some corn and give a basic idea.

    • jrflowers  4 days ago

      The answer is still Absolutely Not, especially since all food can involve a treatment decision for people with type 1 diabetes.

      Pizza is a good example of why not. Slices come in very different sizes, sauces have very different carb content, so do crusts, and toppings.

      Edit: for example this pizza(1) is 31g per slice and this pizza(2) is 73g per slice. The difference is very meaningful and the “general idea” given by photo recognition would likely be wrong to the point of dangerous for a diabetic in both cases.

      If you’re looking for software that can make a guess simply for the sake of generating a number to write down and not be used in any way, a random number generator would be safer since the risk of output being misconstrued as actual information is much lower.

      1 https://www.costcobusinessdelivery.com/kirkland-signature-ca...

      2 https://sbarro.is/product/bbq/

      6 replies →

    • jevogel  4 days ago

      Yeah, you can try this on the ChatGPT app. Take a picture and ask ChatGPT to give you the nutrition info, then do your own calculations based on weight and the USDA database and see how it compares.

      1 reply →

  • je42  4 days ago

    However, if you dont have carb info, the alternative is to judge yourself. Your own model may be better than gpts model, though. I would use GPTs output and at least look at it on a case by case basis