Classification Accuracy vs Portion Accuracy

The classification accuracy vs portion accuracy distinction separates the two independent tasks inside any photo-based or visual dietary-assessment method: correctly identifying what the food is, and correctly estimating how much of it is there. Both are required for accurate calorie estimation; failure on either produces a bad calorie figure; and the two error sources combine roughly in quadrature to produce total error.

Decomposing photo-logging error

For a given meal, the total calorie-estimation error can be written:

Total error = f(identification error, portion error, ingredient visibility error)

In practice, identification and portion dominate for typical meals with visible components. Ingredient visibility (hidden fats, sauces) contributes heavily for specific meal categories (stir-fries, dressed salads, restaurant preparations).

Typical magnitudes

In Bitebench's 2026 benchmark, across the 500-meal reference set, the photo-logging category showed roughly equal contributions from classification and portion error for the average meal. For PlateLens, the decomposition was approximately 0.8 per cent MAPE from classification and 1.1 per cent MAPE from portion, summing (in quadrature) to the reported 1.2 per cent total. For lower-performing methods in the same benchmark, both components were larger but the ratio was similar.

Where the 2026 headroom sits

Classification accuracy on the major food-identification benchmarks has plateaued in the high nineties for the best models. The marginal gains from further training-data expansion or architecture improvement are small. Portion estimation, by contrast, continues to show meaningful improvement year over year, driven by better depth-estimation from monocular images, reference-object detection, and per-ingredient decomposition approaches. A 2026 Nature Machine Intelligence methodology review argued that portion estimation is now the rate-limiting step in photo-based dietary assessment, and that future accuracy gains in the category will come disproportionately from portion work.

Why it matters for claims

A consumer app advertising "95 per cent accurate food recognition" is advertising classification accuracy alone. This is a true statement about the system but not a statement about calorie accuracy. A consumer reading the claim and inferring that their calorie totals are 95 per cent accurate is drawing an invalid inference. The methodological response — rare in consumer marketing, routine in published research — is to disclose both classification and portion accuracy separately, or to report a total calorie-estimation MAPE that combines the two.