ImageNet-D: A New Challenging Robustness Dataset Inspired by Domain Adaptation
Abstract
We propose a new challenging dataset to benchmark robustness of ImageNet-trained models with respect to domain shifts: ImageNet-D. ImageNet- D has six different domains (“Real”, “Painting”, “Clipart”, “Sketch”, “Infograph” and “Quickdraw”). We show that even state-of-the-art models struggle on this dataset and find that they make well-interpretable errors. For example, our best EfficientNet-L2 model experiences a large performance drop even on the “Real” domain from 11.6% on ImageNet clean to 29.2% on the “Real” domain.
Cite
Text
Rusak et al. "ImageNet-D: A New Challenging Robustness Dataset Inspired by Domain Adaptation." ICML 2022 Workshops: Shift_Happens, 2022.Markdown
[Rusak et al. "ImageNet-D: A New Challenging Robustness Dataset Inspired by Domain Adaptation." ICML 2022 Workshops: Shift_Happens, 2022.](https://mlanthology.org/icmlw/2022/rusak2022icmlw-imagenetd/)BibTeX
@inproceedings{rusak2022icmlw-imagenetd,
title = {{ImageNet-D: A New Challenging Robustness Dataset Inspired by Domain Adaptation}},
author = {Rusak, Evgenia and Schneider, Steffen and Gehler, Peter Vincent and Bringmann, Oliver and Brendel, Wieland and Bethge, Matthias},
booktitle = {ICML 2022 Workshops: Shift_Happens},
year = {2022},
url = {https://mlanthology.org/icmlw/2022/rusak2022icmlw-imagenetd/}
}