ImageNet-trained CNNs are biased towards texture (2019)

They are, and it is not that new; the interesting part is how to solve that.

For image classification, this bias for textures may be passable. For image generation, it would produce atrocious results, especially for faces. To the point, that people devised some methods to overcome this, especially with progressive networks (e.g. proGAN). See https://towardsdatascience.com/progan-how-nvidia-generated-i... for the proGAN explanation - it starts from training on 4x4 pixel images, to get the overall shape.

Texture is area-filling, while silhouette is "just a line", so it sounds natural that texture areas would weigh more in an image? If there is both texture and contour "signal" in the labeled inputs, how to pick what get's weighted more? This is supposed to be learned by training.

My favorite titled ml article [1], from 2015, seems to argue for the same conclusion - that it is textures not shapes that drives recognition in CNN

[1] “Suddenly, a leopard print sofa appears”, http://rocknrollnerd.github.io/ml/2015/05/27/leopard-sofa.ht...

I'm curious if stereo vision and eyepoint movement has a benefit for humans in distinguishing shape vs texture, and perhaps that could be a promising approach for ML, but are there any databases for that? There are car focused ones like KITTI, but any more generalist ones?

Quanta article:

https://www.quantamagazine.org/where-we-see-shapes-ai-sees-t...

Style transfer is done using neural networks, so it should be natural that the resultant image fools a neural network classifier. On the other hand, for the filled silhouette, the classifier network was pretty close to human.

What would be a better dataset to train CNNs?