what's the difference of this method to first resizing the feature map to original image size

[ypxie]

tand crop, then resize the cropped feature maps to a fixed size? This can be easily implemented with pytorch functions?