Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

I always just assumed these were terrible.

There just isn’t enough info in a photo, even for the world’s leading calorie counting expert, to give a sensible estimate of the calories content. Ans even more so for macros.

Add to that the fact vision is still pretty crap and you have a complete joke that I can’t believe anyone is paying for.



I've used the AI features in Lose It and was pretty impressed, not for the caloric estimate, but when given a picture of a breakfast burrito it accurately split it out into the component parts, from there I could manually adjust amounts easily for food I make without having to manually search out every ingredient manually. As an additional tool it is great.

The issue of course is these new apps built with a sole focus of AI images for tracking. With a photo of restaurant food you can't see a sauce and the 5-10g of additional sugar content, can't get accurate guesses of what is in breads/pastas and figuring out general volumes of foods is not possible unless you have some kind of standard size reference in the frame.


Honestly the best use case i have seen is the ability to read a labels and add it to the database. (Similar to an OCR but better)

Correcting errors in MFP is the bane of my existence.

I have tried some other apps but i hated every single one of them for a reason or another and came back to MFP but that was one of the best features i would really love to have


What’s interesting to me is a tool can be terrible at its stated problem and still add value.

I suspect many of the results dieters get is “mere measurement”, e.g. it helps them be more thoughtful, seek education, etc.

So the value of the tool is in its ability to keep them engaged and hopeful (the unstated need), and the actual mechanic doesn’t have to be that good at the stated need.


It can also be a negative value.

If I diligently track calories using one of these things and don't lose weight I'll say "screw it" and stop bothering because obviously counting calories doesn't work. I think this is much more likely to be the case because even proper calorie counting is difficult and frustrating


It’s the ‘What gets measured gets managed’ truism - but does it hold true is your measurement is BS?


It’s slightly different- the act of responding itself changes behavior, regardless of any outcome:

https://pmc.ncbi.nlm.nih.gov/articles/PMC12064450/#:~:text=T...


There is a study somewhere that said people who wanted to lose weight and simply tracked their calories (without any attempt at reducing them) LOST weight vs those who did nothing. Awareness alone is valuable.

Maybe this one: https://today.duke.edu/2019/02/tracking-food-leads-losing-po...


I bet @onionweigher could - if all you ate were onions



Yeah. It doesn’t take much imagination to figure out that a Coke Zero looks identical to regular Coca Cola, yet they have vastly different calorie content, to put it mildly.

Sauces loaded with butter, sugar or other goodies will of course be the same story.


However bad these algorithms are (and we're talking very wide intervals in multiple dimensions of uncertainty), you could immediately improve the vision/volume calculation part with a modest change - like requiring a size reference. "Place a unit of currency next to the plate, then take at least two photos of the plate from different angles" can get you pretty close to an accurate volume.

If they aren't doing even that much, then they're not very interested in accuracy at all.


The problem is, if you are serious about losing weight, you need to ideally aim for something like a 300 calorie deficit. Much more and you’ll feel shit, a lot less and it’ll be quite slow or go the other way. It’s quite a fine margin and even tracking using bar codes you’re not going to be perfect.

But even more important - you really gotta track your macros if you want it to work well - and again also not feel crap.

There is no way in hell an AI camera is tracking macros well


The article says that at least the first app trialed does recommend putting a size reference in.


I have faith that ML can approximate calories from images under the following conditions:

- The food is inside containers of a standard size, which the model can use as a volumetric reference

- Optimally, the containers are clear plastic, for depth approximation

- The food is homogenous, or mostly-homogenous. Dishes containing a lot of mixed ingredients seem like an intractable problem to me

Now, whether or not abilities like this are useful to you would seem heavily dependent on your diet.


So here’s why that ain’t true.

Yoghurt.

You can buy zero fat Greek yogurt that has very low sugar. It’s perfect for losing weight. It’s about 50kcals per 100g. And also super high in protein.

Or you can buy yoghurt that is full of sugar and is 100kcals per 100g and has lots of fat and hardly any protein.

Even a human expert could not tell the difference without tasting them.

This is just one example of many.


I wanted to examine this numerically.

* Great Value Original Vanilla Lowfat Yogurt, 32 oz

Serving size 3/4 cup (170g)

1.5g fat

26g carbs (21 of which are sugar)

5g protein

130 kcal

* Great Value Greek Plain Nonfat Yogurt, 32 oz Tub

Serving size 2/3 cup (170g)

7g carbs (7 of which are sugar)

17g protein

100 kcal

* Great Value Light Vanilla Nonfat Yogurt, 32 oz

Serving size 3/4 cup (170g)

15g carbs (12 of which are sugar)

5g protein

80 kcal

If it's only got 50kcal per 100g, then I assume you've got to be relying heavily on indigestible gelling agents to keep the texture reading to the customers as yogurt. I assume that the developer would suggest that a zero-calorie bowl of water and indigestible gelling agents that reads to YOU as yogurt, is not accurately summarized as yogurt, and that this would be a case of user error.


https://www.sainsburys.co.uk/gol-ui/product/fage-total-0-nat...

This is the yoghurt I was talking about. It’s a high quality brand here in the UK

54kal per 100g and 10g protein

3g carbs (zero of which sugar)

Ingredients: Pasteurised Skimmed Milk, Live Active Yoghurt Cultures (L. Bulgaricus, S. Thermophilus, L. Acidophilus, Bifidus, L. Casei)


My mistake. This is very close to the greek yogurt I mentioned, but the serving size is different - I was using 170g recommended serving size instead of the standard 100g comparator. Also - all 3g carbs are sugar (naturally occurring).


And how is it going to tell how much oil or butter was used in a dish?




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: