Your vision model can't weigh food — and that's fine if you architect for it
Описание и применение
Модель распознавания изображений точно определяет плотность пищи, но систематически переоценивает массу, что приводит к неверным калорийным оценкам. Разделив задачу на распознавание, проверенную базу данных и ручное подтверждение порции, команда создала удобный и надёжный поток работы.
Задача
Food tracking apps suffer from friction because users must manually search for foods; vision models can identify food but systematically overestimate portion mass, leading to inaccurate calorie estimates.
Результат
By separating density, mass, and user‑provided portion, the app achieved accurate calorie tracking with minimal user effort, turning a systematic model bias into a trusted workflow.
Рабочий процесс
- 1. Capture photo of meal
- 2. Vision model identifies dish and returns per‑100g nutrition
- 3. System queries database for verified macros of identified dish
- 4. Model provides a mass estimate (biased)
- 5. UI pre‑fills grams with estimate and lets user adjust in one tap
- 6. App calculates total calories and macros from user‑confirmed grams and database values
Доказательства
On mass, they were consistently, directionally wrong. A chicken portion we weighed at 162g came back around 180g. Across the models we tested, overestimates ranged from roughly +10% to well over +100% on identical images.
We ran reference meals with known ground truth (weighed, with composition calculated from reference data) across several models and compared average error.
- Дата записи: 2026-09-08T10:32:00.933Z