Casting Geometric Constraints in Semantic Segmentation as Semi-Supervised Learning