Meta-Dataset is a large-scale few-shot learning benchmark. This dataset does not restrict the few-shot task (it does not require a fixed method or lens), so it represents a more realistic scenario.

The dataset consists of 10 datasets from different fields:

1. ILSVRC-2012 (ImageNet dataset, consisting of natural images from 1,000 categories.)
2. Omniglot (handwritten characters, including 1,623 categories)
3. Aircraft (aircraft image dataset, containing 100 categories)
4. CUB-200-2011 (bird data set, containing 200 categories)
5. Describable Textures (different types of texture images, including 43 categories)
6. Quick Draw (345 black and white sketches in different categories)
7. Fungi (a large mushroom dataset with 1,500 classes)
8. VGG Flower (a dataset of 102 flower images),
9. Traffic Signs (German traffic sign images, including 43 categories)
10. MSCOCO (pictures collected from Flickr, including 80 categories)

Traffic Sign (GTSRB) and COCO datasets in Meta-Dataset are not used for training and are only used for verification or testing. The remaining 8 datasets are divided into training/verification/test sets roughly in the ratio of 70%/15%/15%.