In training deep neural networks for semantic segmentation, the main limiting
factor is the low amount of ground truth annotation data that is available in
currently existing datasets. The limited availability of such data is due to
the time cost and human effort required to accurately