Focal loss is typical for class imbalanced data.

My brain prefers solutions that modify the training objective to resampling based solutions on the grounds of elegance. I'm not sure if that's a good attitude to have or not, if anyone has any thoughts.