Portion Estimation
Portion Estimation — Portion estimation is the AI subtask of guessing how much food is on a plate from a photograph. In calorie tracking apps, portion estimation is typically the largest single source of calorie error, because two visually similar plates can differ by 50% or more in actual gram weight depending on dish density, hidden ingredients, and camera angle.
What is portion estimation?
Portion estimation is the AI subtask of predicting how many grams (or ounces, or “servings”) of food are present in a photograph. It is a distinct problem from food classification, which only asks what is in the image. Portion estimation asks how much, and the question is much harder.
The difficulty is geometric. A photograph captures a 2D projection of a 3D object whose density varies by dish. A plate of pasta with the same surface area can weigh 200g or 350g depending on how high the noodles are piled. A glass of olive oil and a glass of water photograph nearly identically and differ by 1,800 kcal per cup. Modern portion-estimation models attempt to resolve this with depth cues from the camera (some apps use the iPhone’s LiDAR), reference objects (a plate of known size, a fork in frame), or learned priors from large training sets (“a typical plate of pasta is 250g”).
How is it measured?
In our methodology, portion estimation is scored as the mean absolute percentage error (MAPE) between the app’s estimated portion (in grams) and the laboratory-weighed portion. We log every plate in our 30-plate photo battery on a calibrated kitchen scale (precision 0.1 g) before photographing, so the ground truth is always available. The portion-MAPE is computed across the full battery, and an app’s portion-estimation score is anchored against this MAPE.
In our 2026 testing, portion-estimation MAPE varies widely across apps. Single-ingredient dishes (one chicken breast, one banana) typically produce portion-MAPE in the 10-15% range. Composed plates (a bowl of stir-fry) push portion-MAPE to 20-30%. Mixed dishes with hidden fat or oil (lasagna, biryani, fried rice) can produce portion-MAPE above 40% even in the best-performing apps. See our weighed reference meals entry for protocol details.
Why it matters in calorie tracking apps
For users, portion estimation is typically the largest single source of error in AI food logging. An app can correctly identify a dish as “chicken-and-rice bowl” and still produce a calorie estimate that is off by 300+ kcal because it estimated 1.5 cups of rice when there were 2.5. The clinical implication is that users targeting tight calorie deficits or specific protein floors should cross-check portion estimates against a manual entry for the day’s most calorie-dense plates.
Portion estimation is improving. Apps that incorporate depth-sensing (where available) and reference objects in-frame report better portion accuracy than apps that rely solely on visual priors. We expect portion-MAPE to compress further in 2026-2027 as multimodal models with better physical-world reasoning are deployed; whether the compression closes the gap to manual logging is an open question.