We want to use a fine-grained classifier, and it will probably work better with fewer classes to resolve.

The pipeline will work as follows:

• Train a CNN on all classes and cluster the classes into groups which are separable between groups, but confusable within groups.
• Train a Stacked Keypoint Evidence Tree for each on each group to distinguish between their related member classes.
• Offer predictions on test data based on their (probability of being in each group) * (probability of being in each class given membership datum is in this group)

To do this we first need to set up the code for the Stacked Keypoint Evidence Tree.

We may be able to begin with a simpler scheme in which we use superclasses as our groups.