ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness

Geirhos, Robert, Rubisch, Patricia, Michaelis, Claudio · arXiv · 2018