WONT DO: Another tutorial, DeepHomography, use of DataSet API
DeepHomography looks ok to implement. Also, dataset can be installed with tensorflow-dataset. So, it seems a reasonable exercise, as I want to find out if using versioned datasets is worth it.
i.e. is it worth wrapping our datasets using the same mechanism, so people can just pip install, and the API then supports iterating over images, and batching.