nutriingredientnet-v6

NutriIngredientNet v6

Single-RGB-photo food nutrition estimator trained on Nutrition5k (overhead RGB). No depth sensor required at inference — depth is predicted from the RGB image (Depth Anything V2) and fused as a 4th channel, following DPF-Nutrition (Han et al., 2023).

Test results (official RGB test split; T3 real full-dish, comparable to the paper)

• Reconciled calorie MAE: 73.5 kcal (29.1%)
• Blend / Direct / Soft / Hard: 72.7 / 75.3 / 99.1 / 106.1 kcal
• Mass MAE 70.9 g | macro MAE: fat 5.9 g, carb 6.5 g, protein 4.9 g
• Detection F1 0.639 (vs frequency-prior 0.110)
• Grams MAE 20.0 g (vs oracle-presence mean-gram 23.4 g)
• Atwater residual: 14.83 kcal (direct) -> 0.0000 kcal (reconciled, exact)
• Conformal: CRC lambda 0.910 (empirical recall 0.904); gram interval ±49.9 g (empirical coverage 91.1%)

Comparison

• Nutrition5k RGB-only baseline (Thames et al., CVPR 2021): 70.6 kcal / 26.1% (mean 254.94 kcal)
• DPF-Nutrition predicted-depth (Han et al., 2023): 14.7% calorie PMAE
• Prior versions: v2 117.8 | v3 76.5 | v5 77.6 kcal
• Like-for-like published (single/multi image, not 230k frames): VIF2 2025 = 83.75, V-Nutri 2026 = 73.2

Limitations (read before use)

• Fixed overhead camera geometry. Trained on the Nutrition5k rig; other angles degrade accuracy.
• Vocabulary / OOV ceiling. Only ingredients seen >= 10 times in train are modeled; novel foods are missed.
• Extreme-dish under-prediction. Calorie-dense low-volume foods (e.g. 119 g almonds in dish_1558373159, true 1013 kcal) are systematically under-predicted — beyond current published SOTA.
• Conformal caveat. Guarantees assume exchangeability; real-user photos are off-distribution.
• Not a medical device; for behavioral/trend feedback only (Atwater factors are ±5% at best).