Login / Signup
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
Published in:
CoRR (2018)
Keyphrases
</>
texture information
computational efficiency
shape from texture
shape model
multiple scales
image structure
surface texture
classification accuracy
shape prior
surface orientation
video objects
shape representation
training data
shape descriptors
texture analysis
texture features
high accuracy
training set