Embedding-Free Transformer with Inference Spatial Reduction for Efficient Semantic Segmentation