shalev396/calorie-lens-data

Calorie Lens data

Training data of shalev396/calorie-lens: a photo of a meal -> the grams of each of 253 ingredients -> nutrition. Built by training/src/dataset_build.py from:

• Nutrition5k (Thames et al., CVPR 2021, CC BY 4.0): cafeteria dishes weighed ingredient by ingredient, an overhead RGB-D scan and four side-angle videos per dish.
• USDA FoodData Central (public domain, CC0): Foundation Foods, SR Legacy, FNDDS 2021-2023 and Branded Foods, nutrients per 100 g.

Contents

Config	Rows	What
overhead	one per dish with a scan	image (640x480 JPEG q95), depth (uint16 PNG, value / 10000 = metres; 0 = no reading)
side	5 per camera	camera A-D, frame index of n_frames, image (910x512 JPEG q90 from the 1080p video)
dishes	5005	corrected totals (kcal, mass g, fat, carb, protein), the logged ones as orig_*, group, split, test_clean
items	28446	one row per (dish, ingredient): k (model output index), grams, nutrients
ingredients	253	per-gram kcal / fat / carb / protein (Nutrition5k, exact), fiber / sugars / sat_fat / sodium (mg) from the reviewed USDA match
usda	see config	name, category, alternative names, nutrients per 100 g (sodium in mg)

Splits

• test = the official Nutrition5k rgb_test dishes (709); those with an overhead scan are exactly the official depth_test split, so results compare with published numbers. test_clean marks the 565 test dishes whose plate never appears in train / validation.
• validation = 430 dishes: whole plate groups drawn from train. train = 3866 dishes (Cafe 2 included).
• Plate groups: plates were built one item at a time with a scan after each, so many dishes are earlier states of a later one. A group never straddles train and validation.

Changes to the original labels

Every Nutrition5k ingredient's nutrients are grams x a per-gram table; these tables keep that exactly. Fixes (counts):

• Cafe 2 totals logged as 0: 932
• scale total != sum of ingredient grams: 449
• follows the ingredient fixes: 82
• x1000 scale error: 10
• name taken from the table: 8
• empty-plate weight, not food: 2
• photo shows a phone, not the plate: 1
• duplicate of 'watermelon': 1
• duplicate of 'apple': 1

FlavoriaFoodWeight1700 (second cafeteria, weighed plates)

FlavoriaFoodWeight1700 (University of Turku, CC BY 4.0): an automated lunch line photographs each tray from above and weighs every serving it takes. Built by training/src/flavoria.py:

Config	Rows	What
flavoria	856	dish_id, image (the "identification" photo, 1024 px PNG -> 512 x 512 JPEG q90)
flavoria_sessions	856	plate weight mass (g, sum of the weighed servings), date, split, n_items, mapped_g
flavoria_items	per plate	each weighed component, its grams, and the Nutrition5k ingredient (k) where one plain food matches

• Weight only: the source has no calories. 42% of the weighed grams map to a Nutrition5k ingredient (labels/flavoria_component_map.csv, reviewed by hand); mixed dishes count in the plate weight only.
• Splits by lunch day (32 days; every 5th is a test day): train 555, test 301 plates. Median plate 472 g; 138 plates over 600 g.
• Cleaning, of 3345 sessions:
  • soup or dessert (served outside the plate): 95
  • one component over 600 g (scale glitch): 378
  • plate over 1200 g or under 30 g: 384 then 883 identification photos matched to 856 kept plates (tray RFID + time to the second). Sauces from the self-service station and bread are not weighed: weights are slightly low.

Sarapisto, T., Koivunen, L., Mäkilä, T., Klami, A., Ojansivu, P. (2022). FlavoriaFoodWeight1700: Automated Lunch Line Meal Pictures with Automatic Measurement of Weight and Contents. Zenodo. https://doi.org/10.5281/zenodo.5850856

Citation

@inproceedings{thames2021nutrition5k,
  title={Nutrition5k: Towards Automatic Nutritional Understanding of Generic Food},
  author={Thames, Quin and Karpur, Arjun and Norris, Wade and Xia, Fangting and Panait, Liviu and Weyand, Tobias and Sim, Jack},
  booktitle={CVPR}, pages={8903--8911}, year={2021}
}

U.S. Department of Agriculture, Agricultural Research Service. FoodData Central, 2019-2026. fdc.nal.usda.gov.