| Attention is only computed within a local window, and the window is shifted between attention layers to create connections to help the model learn better. Since the Swin Transformer can produce hierarchical feature maps, it is a good candidate for dense prediction tasks like segmentation and detection. |