[LPisGood]

>and even trained professionals are still off by 40%

I find this very hard to believe, unless the term “trained professional” is quite broad. When I was much more into fitness and weighed every meal to the gram, I could tell if a bowl of cereal was a serving to within a gram or two.

[wnorris510]

If you have a known bowl and fill it to a known position every day with the same type of food, then you can probably do better than the average for that specific meal. In our research we've found a majority of calories for most people come from when they're eating out and consuming new dishes where they don't know the ingredients or portion sizes.

In the study we gave people a variety of dishes to make their estimate on, some they were familiar with, some they were not.

The professionals were nutritionists who had trained in portion size estimation and were shown 2D images on a computer screen.

For what it's worth, we've had a lot of people who have claimed to be very accurate at portion size estimation from a long history of using a kitchen scale. We've paid many of them to do a quiz to see if they're above average accuracy and they have almost always ended up around 40% accuracy or worse.

[shawabawa3]

estimating from a photograph is always going to have huge error because you just cannot know e.g. the size of the plate without some external reference

[wnorris510]

Modern phones have LiDAR that give you the near exact volume and dimensions of everything on a plate. That's what we do:

https://www.snapcalorie.com/

We co-authored a paper on this with Google AI and showed it got about 2x the accuracy of a nutritionist because of it.

[maeil]

I'd love to do such a quiz -I might even be willing to pay for the privilege! I'm quite convinced I'm really accurate at calorie estimation without using a scale but would love to be proven wrong. Zero food industry experience here, just from reading hundreds of food labels per year since very young, maybe 8 years old.

Thinking about it again, I'll probably do a lot worse from a picture because I can't have a bite of the food! Just having a spoon makes it so much easier in terms of ratios.

[watwut]

There is no profession that would require you to estimate portion sizes up to grams visually. So, trained professional will be someone who was trained in something different - a doctor for example.

I guess, maybe cooks should have the best precision for this.

[zeroonetwothree]

If you read the paper it’s pretty easy to see what they mean by this. They tested “4 professional nutritionists”. I don’t know if nutritionists get any special training at estimating portion size but my guess would be they do not.

[wnorris510]

Some do, some do not. We put them through a standard portion size training course regardless to be sure.

[ses1984]

Isn’t that a bit of a special case because you know your cereal and you know your bowls? What about some cooked foods like meats which can vary in density and shape when raw, and also vary even further due to inconsistency in cooking, with more or less moisture cooked off?

It’s possible to calibrate your estimates, but if you haven’t done that, it’s probably safe to assume you’re not particularly accurate.

[wnorris510]

There is definitely a lot of variation in density, moisture content, fat percentage between regions, cuts, cooking amounts and methods. IMO using an average number here is probably best because to some extent it's hopeless to account for all of these things.

Most people don't stay consistent in tracking long enough for any of this to matter, so really it's about what is the most accurate approach to achieve your goal and sustain longer term.

[LPisGood]

Oh I would only weigh things raw - if we’re talking about guessing the portion sizes at a restaurant for example, you might say I’m cooked.

I wonder how good an ML model might be at that task. Maybe given a photograph of the plate and the menu description.

[acomms]

I think they're suggesting that the portions you are judging have not been practiced hundreds of times.