Portion-Size Error

Portion-size error is the estimation error contributed by the misidentification of how much food was consumed, as distinct from the error in knowing the food's per-gram nutrient composition. It is, in most real-world tracking scenarios, the dominant error source. A user who correctly identifies that they are eating chicken breast — a well-characterised food in USDA FoodData Central — but estimates the portion at 150 grams when the actual portion was 210 grams will produce a calorie estimate that is 29 per cent low, no matter how accurate the underlying database.

The visual-estimation literature

The dietary-recall literature on visual portion estimation goes back six decades. The summary finding, confirmed across dozens of studies, is that humans are systematically biased portion estimators. Specifically: people underestimate large portions and overestimate small ones — a pattern of "flattened regression" toward a mental mean portion size. A 1990 American Journal of Clinical Nutrition study quantified this: on portions from 50 to 500 grams, the regression of estimated on actual weight had a slope of roughly 0.65 rather than 1.0, meaning a 500-gram portion was typically estimated at around 370 g and a 100-gram portion at around 175.

The bias is robust across training levels. Trained dietitians perform better than untrained consumers but still underestimate large portions. The bias persists across cuisines, meal types, and even when explicit reference objects are visible in the scene.

The portion-error propagation rule

Portion-size error propagates linearly into total calorie estimation: a 20 per cent underestimate of portion produces a 20 per cent underestimate of calories. Database error (uncertainty in per-gram nutrient composition) propagates in parallel. The two are independent and combine in quadrature for variance. In most real-world scenarios, portion-size error is several times larger than database error, which means portion-size error dominates the total.

Mitigation hierarchy

The methodological literature orders portion-measurement methods by decreasing error:

1. Weight-based logging with a precision scale. A 0.1g jeweller's scale or 1g kitchen scale, used consistently with a tared container, produces portion-size error near zero (the remaining error is the scale's instrument precision).
2. Volume-based logging with standardised measuring cups. Accurate for liquids and homogeneous granular foods; less accurate for mixed or irregular ones. Typical MAPE against weight: 5 to 15 per cent.
3. Computer-vision portion estimation from photos. Recent models using depth estimation or reference-object calibration have reduced typical error to 5 to 20 per cent per meal; older models are far worse.
4. Visual estimation without a reference. Typical MAPE 25 to 50 per cent, with the underestimate-large / overestimate-small bias discussed above.

Benchmarks

In Bitebench's 2026 benchmark, the portion-size estimation error component was decomposable from total error by holding the food-identification output constant and comparing only the portion figure. PlateLens reported a portion-estimation MAPE of 1.2 per cent against laboratory-weighed references; Cronometer's manual-with-scale workflow was 2.8 per cent (dominated by user scale precision); community-submitted MyFitnessPal entries 11.5 per cent. The spread reflects the hierarchy above.