ImageNet trained CNNs are biased towards recognising textures rather than shapes

It seems to me that CNNs like that are unlikely to really understand the 3D structure of objects.

To me the ideal vision system would know about surfaces and shapes made from surfaces as well as shapes made from smaller parts. It would be able to use clues from the textures such as writing on a can to detect surfaces but also go the other way and given a hypothesis about a surface shape, use that to help read the writing.

Anyone know of a system like that? Would it use CNNs, or something else, like capsules?

But a CNN like that never sees any object in 3D, never a stereo pair, usually not even the same object from different aspects (marked as the same object). There is no information from which an idea of the 3D structure could arise.

Wasn't it the other way around? That cow fitted into a shark body shows shape being the key aspect.

https://chriswolfvision.medium.com/cow-sharks-exploring-the-...

Isn't it already known?

Yes but this research explains better how this happens. And it’s from 2019. I thought it’s an interesting paper.