DeMT: Deformable Mixer Transformer for Multi-Task Learning of Dense Prediction