Re:Where we see shapes, AI sees textures
by hammock on Monday July 08, 2019 @02:43PM
(#20385509)
I wonder what would happen if we fed it SVGs instead of JPGs. (Shape files)
[ Parent
| on HN ]
At that point it's less a task in the context of image processing and more one of graph learning / language processing, depending on how you want to formulate it.
Human vision certainly contains a component where we do 3d reconstruction and adjust our perception based on that. Many optical illusions depend on this step.
If we want more human like machine vision then having passes in image processing that deal with more abstract data sounds like a great idea
Re:Where we see shapes, AI sees textures
by mirekrusin on Monday July 08, 2019 @02:54PM
(#20385622)
I don't think those networks can work with variable inputs, they use fixed size images, svg describes paths that highly vary.
That's true for convolutional networks. However, recurrent architectures aren't similarly limited, e.g SketchRNN:
https://github.com/tensorflow/magenta/blob/master/magenta/mo...