I have an issue, where I need to compare shapes that appear in images.
The images might show the same thing, but their content might be twisted, turned, ... a conventional algorithm would therefore fail.
There shapes that might appear are unlimited, i need a comparison based on the concept, not their representing class.
I've tried Siamese networks with Tensorflow, but with really bad results. It either didn't learn scored images way too closed or wasn't able to compare images from the testing data.
