How accurate is a photo calorie estimate?

Every app that photographs food makes the same implicit promise: point, shoot, know. It is a good promise, and it is partly true. The part that is not true is usually left unsaid, which is how people end up trusting a number they should have questioned.

So here is the honest version.

What the camera is actually good at

Two things, and it is genuinely good at both.

Identifying what is on the plate. Modern food recognition is strong at naming things. A bowl with grilled chicken, brown rice, black beans, guacamole and corn salsa comes back as five separate items, not one guess at "burrito bowl". That matters more than it sounds: the biggest source of error in food logging has never been arithmetic, it is people not logging at all because typing five entries is tedious.

Being consistent. If you photograph the same lunch on Monday and Thursday, you get roughly the same answer. Consistency is what makes a trend readable, even when the absolute number is off. A log that is uniformly 10% high still shows you the week you ate more than usual.

What it is bad at

Portion size. This is the hard one, and it is hard for physical reasons rather than fixable ones. A photograph is a flat projection of a three-dimensional thing. Depth is inferred, not measured. A bowl of rice photographed from above looks much the same whether it is three centimetres deep or eight.

What it cannot see. Oil absorbed into fried food. Sugar dissolved in a sauce. The butter a dish was finished with. Two visually identical plates of fried rice can differ by a couple of hundred calories depending entirely on things that left no trace in the image.

Dishes it has not met. Recognition quality tracks how well represented a cuisine is in the data behind it. This is why a tracker built around one food culture quietly degrades everywhere else. It is not that it refuses to recognise your food; it is that it confidently returns something adjacent.

What the error actually looks like

The shape of the error follows from the physics above: identification is the strong part, quantification is the weak part, and errors on individual items can be large even when the daily total lands close. Two mistakes in opposite directions often cancel.

That cancellation is the useful insight. It means:

• A single meal's number is a rough estimate. Do not build a decision on it.
• A week's numbers are a usable signal. Errors average out; patterns survive.
• A month's numbers are genuinely informative. This is where tracking earns its keep.

What we do about it in Senumi

Three things, none of them magic.

1. You confirm before anything is logged. The scan produces a list of items with portions, and nothing enters your history until you have looked at it. Untick what is not there, adjust a portion that is obviously off.
2. Estimates are labelled as estimates. Where a number is inferred, the app says so. We would rather be visibly uncertain than quietly wrong.
3. Your correction is what gets logged. Fix a misidentified dish and the corrected version, not the guess, is what enters your history.

The useful way to think about it

A photo estimate is not a kitchen scale, and it is not trying to be. It is closer to a step counter: nobody believes their phone counted exactly 9,193 steps, and nobody needs it to. The number is useful because it is the same kind of wrong every day, which makes the shape of the week real even when the digits are not.

Use it to notice that you eat 600 more calories on days you skip breakfast. Do not use it to decide a dose.