Food-101

Overview

The Food-101 dataset is a large-scale image classification benchmark for single-label food recognition. It contains 101 food categories with 1,000 images per class, resulting in 101,000 images in total.

The dataset is intentionally designed to reflect real-world conditions. While the test set is manually reviewed and cleaned, the training set is left noisy, containing mislabeled samples and images with strong color distortions. This makes Food-101 particularly suitable for evaluating robustness and generalization in visual recognition models.

Train: 75,750 images (750 per class, noisy)
Test: 25,250 images (250 per class, manually reviewed)

All images are RGB and were rescaled so that the maximum side length does not exceed 512 pixels, while preserving the original aspect ratio.

Data Structure

When accessing an example using ds[i], you will receive a dictionary with the following keys:

Key	Type	Description
`image`	`PIL.Image.Image`	RGB image with maximum side length ≤ 512 pixels
`label`	int	Class label in the range `[0, 100]`

Usage Example

Basic Usage

from stable_datasets.images.food101 import Food101

# Load training split
ds = Food101(split="train")

# Load all splits as a DatasetDict
ds_all = Food101(split=None)

sample = ds[0]
print(sample.keys())  # {"image", "label"}

# Optional: make it PyTorch-friendly
ds_torch = ds.with_format("torch")

References

Official website: https://data.vision.ee.ethz.ch/cvl/datasets_extra/food-101/
License: Research use (as provided by the original authors)

Citation

@inproceedings{bossard2014food,
  title     = {Food-101 -- Mining Discriminative Components with Random Forests},
  author    = {Bossard, Lukas and Guillaumin, Matthieu and Van Gool, Luc},
  booktitle = {European Conference on Computer Vision},
  year      = {2014}
}