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

Devs

ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness | Read Paper on Bytez