ODL geometry interfaces

[adler-j]

Hello! I just found out about this library and I'm very optimistic about it.

So far most implementations have used something like the as_tensorflow_layer in ODL which calls ASTRA as a backend to get CUDA accelerated projections.

ODL already has rather thought out geometry objects that should be convertible to PYRO-NN geometries. This would allow seamless testing and integration between the framework, e.g. doing a TV reconstruction in ODL and training with PYRO-NN. It would also allow us to validate each others implementations and perform benchmarks.

[csyben]

Hello,
i was planning to contact you either way :).

I think that's a great idea, we definitely should do that. It should be pretty straight forward to bring the PRYO-NN-Layers to ODL. I think we have to ensure that we find all minor differences, or implicit assumptions (e.g. For the 3D projection matrices we assume a normalization such that the last entry is 1, also the channel handling could be a minor obstacle, etc.)