Mask Propagation for Efficient Video Semantic Segmentation