README.md

metadata

license: mit
pretty_name: Control+Shift
size_categories:
  - 10K<n<100K
task_categories:
  - image-classification
dataset_info:
  features:
    - name: shift_type
      dtype: string
    - name: shift_amount
      dtype: int8
    - name: image
      dtype: image
    - name: label
      dtype: int8
    - name: dataset
      dtype: string

Dataset Card for Control+Shift: Generating Controllable Distribution Shifts

[arXiv], [GitHub]

Curated by: Roy Friedman and Rhea Chowers

This dataset is the one that accompanies the paper Control+Shift: Generating Controllable Distribution Shifts. Our data is based on CIFAR10 and ImageNet, using EDM to generate our data. We generated datasets for 3 types of distribution shift on CIFAR10 and ImageNet - so a total of 6 datasets. The types of distribution shifts are called overlap, extend and truncation.