Question about feature map

[ManiadisG]

Hello,

I have a question about the feature map that is extracted by the Swin backbone. Assuming an input with size (224,224), the original Swin model produces 4 feature maps, with shapes (C, 56, 56), (2C, 28, 28), (4C, 14, 14) and (8C, 7, 7).

Your version, however, produces 4 feature maps (2C, 28, 28), (4C, 14, 14), (8C, 7, 7) and (256, 4, 4).

Can you please explain why you are not also using the 1st feature map?