Add Figure Of Merits

[zickert]

Add implementation of figure of merits to quantitatively compare reconstructions with known ground truth. Separate branch with corresponding pull request in preparation.

[adler-j]

A general comment is that if we make a odl.contrib.fom package (which i think we should) then there are some sub-modules that should be kept separate (in my opinion). Some suggested such sub-modules would be (names are obviously preliminary):

• unsupervised: FOMs of the form fom(reconstruction) attempting to estimate features from a single image.
• supervised: FOMs of the form fom(reconstruction, true) computing some generalized distance
• ensamble: FOMs of the form fom(reconstruction1, reconstruction2, ..., reconstructionn, [true]) computing some kind of batch statistic (e.g. noise sensitivity, etc)

[adler-j]

I was thinking about this when implementing some other stuff and I realized that there are some features that would be nice to have. These do NOT need to be implemented in the PR, but should be added shortly after.

• More FOM, in particular we want "signal to noise ratio" (SNR) and "peak signal to noise ratio" (PSNR)
• Utilities to normalize values, either as a parameter to the FOMs (e.g. normalization=...) or by external functions. Examples include z-scores and rescaling into [0, 1].
• Unsupervised FOMs, in particular some type of noise estimation similar to this one

Anyway, would be great to get the main skeleton in so we can start building on top of it.

[adler-j]

Closed by #1018