Multi-Scale High-Resolution Vision Transformer for Semantic Segmentation