New paper on shape vs texture in how NN works

(matej) #1

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

Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann & Wieland Brendel."

Blog on the topic

Idea: implementing shape style transfer to transforms to improve accuracy in fastai

0 Likes