NutriIngredientNet v3
Itemised, energy-consistent nutrition from one overhead RGB photo (Nutrition5k, official RGB split).
v3 adds a direct totals head and reconciles the itemised breakdown to it, buying top-down
accuracy without giving up the audit trail. Scaling all grams by one factor preserves the Atwater
identity exactly (verified numerically in §12).
Test results (official RGB split, calibrated T=1.02, thr=0.68, TTA=True)
- Detection: P 0.61 / R 0.62 / F1 0.62 (freq-prior F1 0.22)
- Grams MAE: 18.7 g (oracle-presence mean-gram baseline 23.5 g)
- Per-ingredient kcal MAE (reconciled, TPs): 25.8 kcal
- Dish kcal MAE by estimator (vs T2 real vocab-covered):
hard 118.3
soft 77.1
direct 84.5
recon 83.8
- Best estimator: soft | T3 (paper-comparable) 76.5 kcal
- reference: Nutrition5k RGB-only 2D-direct baseline = 70.6 kcal (26.1%)
- previous version (v2, bottom-up only) = 117.8 kcal (46.6%)
- OOV ceiling: 7.7 kcal | cost of exact energy consistency: +0.0 kcal
- Conformal set (lambda=0.25): target recall >= 90%, empirical 0.848,
avg set size 11.6
- Gram intervals: +/- 0.81 x mean_gram, empirical coverage 0.876
Honest limitations
- Fixed camera geometry. Nutrition5k's camera height never changes, which is the only reason
apparent size maps to grams. On an uncalibrated phone photo, scale is unrecoverable from a single
view — this model is NOT validated for in-the-wild phone photos.
- Nutrition = grams x a fixed per-gram table; the network does detection + portion only. That is
what buys exact energy consistency and auditability.
- 164-ingredient vocabulary; out-of-vocab food is invisible (8 kcal ceiling).
- Conformal guarantees assume val/test exchangeability, which the official split does not perfectly
satisfy; empirical numbers above.
- Depth-as-privileged-information head: enabled (deleted at inference).
