I came across the coolest use-case for Blender the other day, made by Andrew Moffat. It turns out that you can use the Python API inside of the 3D tool to generate training data for receipts!
Here are some of the examples that it was able to generate.
It works by using a flat .png file as input, which is
mapped unto a 3D surface that resembles wrinkles.
This is a pretty clever way to generate training data. You can reliably add/remove wrinkles, change the background, add blur and exchange the underlying receipts too.
Very cool to see Blender still getting some attention from the ML community.