Notes on hackmd: https://hackmd.io/rJPX1LAJTsu5LFbdbW0YYg

From meeting (11/01/2023), plan is to try an alternative approach using context in post-processing rather than during training/classification.

Consider examples (X = railspace, o = no):

^ This is likely a false positive

^ This is likely correct

^ This likely requires attention/review

This kind of post-processing can be done fairly easily using rules to pick out which ones we want to change labels for and/or review.
See notes in hackmd for more info.

Things to consider:

• How does this generalise across other MapReader use cases? e.g. non-linear features such as buildings or trees? Any post-processing that is included into MapReader code should allow users to define their own rules for false-positives etc. to allow for different use cases

Have just merged #350 which adds functionality for using context images when training a model. Would be good to test this a bit more at some point but am closing this ticket for now.