Try a different single model

[thomashopkins32]

UNet apparently isn't meant for instance segmentation but semantic segmentation (I have to double check that this is accurate).

I should try a different model that was built for instance segmentation once I have squeezed out performance on UNet to the best of my ability.

[thomashopkins32]

Double-checked and it is accurate. The problem with using mAP as a metric with the UNet architecture is that I have to run a connected components analysis which is very expensive. Since I have a limited training budget (and the competition is over...) I can use a simpler metric like mean IoU over the whole image. This should speed up validation significantly and allow me to train for longer.