question about dataset

[ccw12345]

I am confused about the dataset of LEVIR_CD. Do you crop each image directly into 16 non-overlapping 256*256 images？If you do this, there will be many label images that are completely unchanged. How did you deal with it？

[wzjialang]

I use the entire 256*256 image patches as my dataset.

Actually, it is a normal phenomenon that an area does not experience any interesting changes over a period of time. Therefore, from a practical point of view, it is not necessary to filter those unchanged image patches. On the other hand, maybe filtering could lead to better performance due to the characteristics of deep learning (less noise).