It could be really to generate synthetic data for semantic segmentation tasks.