Food-101: 101,000 Images Across 101 Food Categories

ETH Zurich's Food-101 contains 101,000 images distributed across 101 distinct food categories. Each class includes 750 training images and 250 manually reviewed test images, with all files rescaled to a maximum side length of 512 pixels.

Use Cases

• Multi-class image classification using the 101 food category labels
• Training models to handle label noise and color distortion in food photography
• Fine-grained visual recognition of specific dishes like sushi, pizza, or pad thai

Strengths

• 101,000 total images
• 101 distinct food categories
• Manually reviewed test set of 25,250 images
• Standardized 512px maximum side length

Limitations

• Intentional label noise in the training set
• Intense color distortions in some images
• Geographic bias toward US-centric food types as indicated by source tags

Provenance

Source

ETH Zurich (ETHZ) and Foodspotting

Collection Method

Crowdsourced images from Foodspotting with manual review for the test set

Geography

United States