arunapb/food-density-nutrition

FoodSeg103 Food Density & Nutrition Dataset

Dataset Summary

This dataset provides physical density and nutritional information for all 103 food ingredient classes defined in the FoodSeg103 benchmark — a large-scale food image segmentation dataset introduced in the ACM MM 2021 paper "A Large-Scale Benchmark for Food Image Segmentation" (Wu et al., 2021).

The goal is to enrich food segmentation models with physical and nutritional context, enabling downstream tasks such as calorie estimation, portion size calculation, and nutrient profiling from food images.

---

What is FoodSeg103?

FoodSeg103 is a large-scale benchmark for fine-grained food image segmentation containing 7,118 images annotated with 103 ingredient classes and pixel-wise segmentation masks. Each image contains an average of 6 ingredient labels. The dataset was curated from Recipe1M and annotated by human annotators.

• Paper: A Large-Scale Benchmark for Food Image Segmentation
• Authors: Xiongwei Wu, Xin Fu, Ying Liu, Ee-Peng Lim, Steven C.H. Hoi, Qianru Sun
• GitHub: LARC-CMU-SMU/FoodSeg103-Benchmark-v1
• License: Apache 2.0

---

This Dataset

For each of the 103 FoodSeg103 food classes, this dataset provides:

Physical Density (from aqua-calc.com / USDA)

Bulk/apparent density measured as grams per 250 mL metric cup, converted to standard units. Useful for estimating food mass from volume (e.g., from segmentation masks with depth estimation).

Nutritional Information (per 100g, from FatSecret)

18 nutrition columns sourced from FatSecret, which aggregates USDA FoodData Central data.

---

Dataset Structure

Columns

Column	Type	Description
class_id	int	FoodSeg103 class ID (0–102)
class_name	string	FoodSeg103 class label
food_description	string	USDA/aqua-calc food name for density lookup
g_per_metric_cup	float	Grams per 250 mL metric cup (bulk density raw)
density_g_per_cm3	float	Density in g/cm³
density_kg_per_m3	float	Density in kg/m³
calories_kcal	float	Energy (kcal per 100g)
fat_total_g	float	Total fat (g per 100g)
fat_saturated_g	float	Saturated fat (g per 100g)
fat_polyunsat_g	float	Polyunsaturated fat (g per 100g)
fat_monounsat_g	float	Monounsaturated fat (g per 100g)
fat_trans_g	float	Trans fat (g per 100g)
cholesterol_mg	float	Cholesterol (mg per 100g)
sodium_mg	float	Sodium (mg per 100g)
carb_g	float	Total carbohydrates (g per 100g)
fiber_g	float	Dietary fiber (g per 100g)
sugar_g	float	Total sugars (g per 100g)
protein_g	float	Protein (g per 100g)
vitamin_a_ug	float	Vitamin A (µg RAE per 100g)
vitamin_c_mg	float	Vitamin C (mg per 100g)
vitamin_d_ug	float	Vitamin D (µg per 100g)
calcium_mg	float	Calcium (mg per 100g)
iron_mg	float	Iron (mg per 100g)
potassium_mg	float	Potassium (mg per 100g)

---

Data Collection

Density Data

• Source: aqua-calc.com (powered by USDA FoodData Central)
• Method: Web scraping — each food matched to closest USDA entry
• Unit: Bulk/apparent density (includes air gaps), not true solid density
• Conversion: density (g/cm³) = g_per_metric_cup / 250

Nutrition Data

• Source: FatSecret (aggregates USDA FoodData Central / SR Legacy)
• Method: Web scraping — searched by class name, fetched 100g serving page
• Unit: Per 100g of food (consistent across all classes)

---

Usage Examples

from datasets import load_dataset

ds = load_dataset("arunapb/food-density-nutrition", split="train")
df = ds.to_pandas()

# Estimate mass from segmented volume (cm³)
def estimate_mass(class_name, volume_cm3):
    row = df[df["class_name"] == class_name].iloc[0]
    return volume_cm3 * row["density_g_per_cm3"]

# Estimate calories from mass
def estimate_calories(class_name, mass_g):
    row = df[df["class_name"] == class_name].iloc[0]
    return (mass_g / 100) * row["calories_kcal"]

# Example: 200cm³ of apple
mass = estimate_mass("apple", 200)       # ~105.6g
cals = estimate_calories("apple", mass)  # ~54.9 kcal
print(f"Mass: {mass:.1f}g | Calories: {cals:.1f} kcal")

---

Intended Use Cases

• Calorie estimation from food segmentation masks (volume → mass → calories)
• Nutritional profiling of meals from food images
• Portion size analysis using physical density and segmented area
• Dataset enrichment — adding nutritional metadata to FoodSeg103 predictions
• Food computing research — linking vision models to nutritional knowledge

---

Limitations

• Density values are bulk/apparent densities and vary with food preparation and packing
• Nutrition values are generic averages — actual values vary by variety, preparation, and ripeness
• Some classes (e.g. other ingredients, hanamaki baozi) are broad and have approximate values

---

Citation

@inproceedings{wu2021foodseg103,
  title     = {A Large-Scale Benchmark for Food Image Segmentation},
  author    = {Xiongwei Wu and Xin Fu and Ying Liu and Ee-Peng Lim and Steven C.H. Hoi and Qianru Sun},
  booktitle = {Proceedings of the 29th ACM International Conference on Multimedia},
  year      = {2021},
  doi       = {10.1145/3474085.3475628}
}

---

License

• This dataset: CC BY 4.0
• FoodSeg103 class list: Apache 2.0 (Wu et al., 2021)
• Nutritional & density data: USDA FoodData Central (public domain)