arunapb/nutriingredientnet-v2

NutriIngredientNet v2

Per-ingredient food analysis from one overhead RGB photo (Nutrition5k, official RGB split): detects ingredients, estimates grams with calibrated +/- intervals, and outputs an energy-consistent kcal/macro breakdown (kcal = 9fat + 4carb + 4*protein holds exactly, by Atwater projection of the per-gram table; enforcing it cost -2.5 kcal MAE).

Test results (official split, threshold 0.85)

• Detection: P 0.66 / R 0.65 / F1 0.66 (frequency-prior baseline F1 0.22)
• Grams MAE (present): 19.1 g (oracle-presence mean-gram baseline 23.5 g)
• Per-ingredient kcal MAE (true positives, real per-row cal): 25.2 kcal
• Dish kcal MAE — three targets: T1 table-reconstructed 120.7 | T2 real vocab-covered 112.8 | T3 real full dish 117.8 (paper-comparable; OOV ceiling 7.7)
• Conformal set (lambda=0.69): guaranteed recall >= 90% on exchangeable data; empirical test recall 0.806; avg set size 7.2
• Gram intervals: +/- 0.82 x mean_gram; empirical test coverage 0.872 (nominal 0.90)

Honest limitations

• Nutrition = predicted grams x fixed per-gram table (variant A): the network only does detection + portion estimation; nutrition itself is deterministic given grams. This is what buys exact energy consistency.
• 164-ingredient vocabulary; out-of-vocab food is invisible -> the 8 kcal OOV ceiling no training can remove.
• Conformal guarantees assume val/test exchangeability; the official split is not perfectly exchangeable with our val pool — empirical checks above quantify the gap.
• Single overhead RGB view, fixed camera height; not validated on phone photos in the wild.