metadata
license: cc-by-4.0
pretty_name: ImageNet-DomainNet-NvR
This dataset supplies new classification labels to existing image classification/distribution shift robustness benchmarks. The classification task involves distinguishing images with a natural style (e.g., photographs) from renditions (e.g., sketches, and cartoons). There are three classes: "natural," "ambiguous," and "rendition." We provide 1000 labels for each of the seven common benchmark datasets.
These samples should be used only to test new rendition classifiers and not to train/design new ones!
When using this dataset, please also cite the work introducing each of the used data sources (see README.md in each .tar file).