Learning-based approaches for semantic segmentation have two inherent
challenges. First, acquiring pixel-wise labels is expensive and time-consuming.
Second, realistic segmentation datasets are highly unbalanced: some categories
are much more abundant than others, biasing the performan