issai/nutrition-workshop-sample

Nutrition Workshop Sample

A small food object detection set for teaching. It is a sample of the Food Portion Benchmark (FPB), cut down to 20 dishes so that a beginner can train a working detector on a free Colab GPU in under ten minutes.

Built for the Nutrition and AI workshop at ISSAI, Nazarbayev University, for participants with no programming background.

What is in it

Task	object detection, YOLO format
Dishes	20
Photos	1,127
Boxes	1,699
Multi-dish photos	402
Image size	longest side 640 px

Splits are 780 train, 169 validation, 178 test photos.

Dishes

chak_chak, bukteme, kurt, balqaymaq, hvorost, black_tea, fresh_salad, borsh, doner, mayonnaise, nuggets, cheese_sticks, fanta, cupcake, capuccino, croissant, udon, pizza, french_fries, ketchup

The classes were chosen together, not one by one, so that whole cafeteria trays are fully annotated. That is what lets a workshop go from "name this dish" to "add up the calories of this meal". Traditional Kazakh items sit alongside the canteen staples they are actually served with.

Layout

FPB_mini/
  data.yaml
  images/{train,val,test}/*.jpg
  labels/{train,val,test}/*.txt     # class_id x_center y_center width height, normalised

Use it

import urllib.request, zipfile
url = "https://huggingface.co/datasets/issai/nutrition-workshop-sample/resolve/main/FPB_mini.zip"
urllib.request.urlretrieve(url, "FPB_mini.zip")
zipfile.ZipFile("FPB_mini.zip").extractall(".")

from ultralytics import YOLO
YOLO("yolo11n.pt").train(data="FPB_mini/data.yaml", epochs=40, imgsz=640)

How the sample was drawn

Two rules keep this sample honest, and both matter if you compare numbers.

The FPB test split was never read. FPB withholds its test labels so that its public leaderboard stays fair. Every image here comes from the FPB train and validation splits only, so nothing in this sample discloses a hidden test label.

Every dish in every photo is labelled. Photos are kept only when all of their annotated food falls inside these 20 classes. A photo with unlabelled food on the tray would teach a detector that food is background, and would make a "total up the meal" demo quietly wrong.

Splits are separated by physical plate, not by photo. In FPB each portion was photographed from several angles. Splitting at the level of individual photos puts near-duplicate views of the same plate into both train and test, which inflates scores badly. Here all views of one plate stay inside one split.

Scores are therefore lower than a naive split would give, and closer to what the model would do on a new meal.

Limitations

Laboratory photos on canteen trays, so a model trained here will not transfer well to restaurant or home kitchen scenes. Each photo is cropped to the food to remove the rig and the room, but colour-calibration cards and rulers sit on the trays themselves and remain visible in some frames. Portion weights from FPB are not included, only boxes and dish names. Teaching resource, not a clinical tool.

Licence and credit

cc-by-nc-4.0, inherited from the parent dataset. Non-commercial use only.

Please cite the original FPB paper:

@article{fpb2025,
  author  = {Sanatbyek, Aibota and Rakhimzhanova, Tomiris and Nurmanova, Bibinur
             and Omarova, Zhuldyz and Rakhmankulova, Aidana and Orazbayev, Rustem
             and Varol, Huseyin Atakan and Chan, Mei Yen},
  journal = {IEEE Access},
  title   = {A Multitask Deep Learning Model for Food Scene Recognition and
             Portion Estimation - the Food Portion Benchmark (FPB) Dataset},
  year    = {2025},
  volume  = {13},
  pages   = {152033--152045},
  doi     = {10.1109/ACCESS.2025.3603287}
}