Stylized ImageNet 数据集是通过剥离 ImageNet 中的局部纹理，同时通过 AdaIN style transfer 保留自然图像上的全局形状信息而创建的。这促使 CNN 更多地学习形状，减少对局部纹理的学习。

Citation

@inproceedings{geirhos2019, title={ImageNet-trained {CNN}s are biased towards texture; increasing shape bias improves accuracy and robustness.}, author={Robert Geirhos and Patricia Rubisch and Claudio Michaelis and Matthias Bethge and Felix A Wichmann and Wieland Brendel}, booktitle={International Conference on Learning Representations}, year={2019}, url={https://openreview.net/forum?id=Bygh9j09KX}, }