The image URIs with annotations can be downloaded here (20MB). Below explains the file structure in the zip folder.

• mini-imagenet-annotations.json ‐ JSON annotations on Mini-ImageNet
• stanford-cars-annotations.json ‐ JSON annotations on Stanford Cars
• mini-imagenet/
  • split/ - training splits for two noise types and ten noise levels. The two columns are image_id and class_id.
  • class_names.txt - all the class names. The class id is equivalent to (class line number - 1)
• stanford_cars/
  • split/ - training splits for two noise types and ten noise levels.
  • class_names.txt - all the class names. The class id is equivalent to (class line number - 1)

You can find the starter code using our training code.

Please contact us via controlled-noisy-labels@googlegroups.com, if you have any questions or difficulty in reproducing our results on this dataset.