Confusion about input size and output size and center crop?

[WNZhao1988]

Based on your code, it seems that you crop the center from a 540x540 image to get the 270x270 image as input, and for annotation, you crop the center from a 540x540 annotation to get the 80x80 annotation as output/ground truth. I wonder why for input you only use the center part instead od the original size of the image and also for annotation why you don't keep it consistent with the input size (e.g. 270x270 or 540x540), instead, you use a smaller outputsize?

Could you simply explain this to me?

Thank you!

[simongraham]

We start with a patch double the size of the input.

1. 540 x 540 patch
2. augmentation
3. central crop: input size = 270x270
4. output size of network = 80x80

We use an initial image double the size of the input before augmentation to reduce border effects during augmentation (for example, rotation ~ 45 degrees)