Accuracy

How accurate are AI calorie counters?

A plain account of what we measured, why photos are hard to read, and what you can do to get closer.

What we measured

We benchmarked SnapMacro's photo model on 200 meals from the public Nutrition5k dataset. Nutrition5k is a Google Research dataset of real cafeteria dishes in which every ingredient was weighed as it was added to the plate, so the calorie and macro values are measured rather than guessed.¹ That makes it a fair yardstick: the answer key does not depend on anyone's eyesight.

Test	Result
Photo model, 200 Nutrition5k meals, calories	Median error roughly 27 to 33 percent across runs
Direction of error	A tendency to overestimate, about 21 percent on average
Depth study, same dataset: volume estimated from depth data, referenced to the plate	About 26 percent median error
Depth study, same subset: photo model	About 35 percent median error

Two caveats. The test meals are cafeteria plates photographed under controlled conditions, which are not the same as a dinner at home, a restaurant or a packed lunch. And the depth comparison was a study on one subset of one dataset, not a feature of the app; SnapMacro uses an ordinary photo. We share it because it shows where the difficulty lies.

Why photos are hard

Hidden oil and sauce

A tablespoon of oil adds roughly 120 calories and is nearly invisible once it is absorbed into rice, vegetables or meat. Dressings, butter and sauces behave the same way. A camera sees the surface of the food, not what it was cooked in.

Portion depth

A photograph is flat. A shallow plate of pasta and a heaped one can look alike from above, yet differ by hundreds of calories. In the Nutrition5k paper the authors point to depth information as a way to improve predictions, and our own depth comparison above points the same way: estimating volume is the hardest part of the job.

Hidden ingredients

Layers, fillings and mixed dishes conceal what is inside. A burrito, a casserole and a smoothie are all mostly unseen.

What SnapMacro does about it

• Editable items. Every item, portion and the dish name can be changed after a scan.
• Swap suggestions. If the model picked the wrong food, you can swap it for a close alternative instead of starting over.
• Add oil or sauce. You can add the cooking fat or sauce the camera could not see.
• A confidence note. When portions or ingredients are partly a guess, the result says so and asks you to check it.
• Barcode for packaged food. When food has a label, scanning the barcode is more reliable than any photo.

The aim is not a perfect number at every meal. It is a record that is close enough, over enough days, for your weight trend to tell the truth.

How to get a better estimate

• Shoot from about 45 degrees. An angle shows the height of the food as well as its spread, which a straight-down photo hides.
• Include the whole plate. The plate or bowl gives the model a size reference. Do not crop it.
• Mention oil and sauce. If you cooked with oil or added dressing, say so in the note or add it after the scan.
• Separate the components when you can. A plate where foods do not overlap is easier to read than a pile.
• Use the barcode for anything with a label.
• Look at the weekly average. Errors in both directions tend to partly cancel across many meals, which is why a seven-day view is more trustworthy than any single log.

If you need to know a number precisely, for example for a medical reason, weigh the food on a scale and ask your clinician or dietitian. SnapMacro is not a medical device.

Related

Your starting targets come from standard formulas you can try yourself: the TDEE calculator and the protein calculator. See also how we compare with other trackers: SnapMacro vs Cal AI and SnapMacro vs MacroFactor.

Source

1. Thames Q, Karpur A, Norris W, Xia F, Panait L, Weyand T, Sim J. Nutrition5k: Towards Automatic Nutritional Understanding of Generic Food. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. openaccess.thecvf.com