search

llu025 / Heterogeneous_CD

This repository contains the software developed for the project of Heterogeneous Change Detection in Remote Sensing Images.
107 stars 39 forks source link

about the final_evaluate() #13

Closed haoranwu22 closed 2 years ago

haoranwu22 commented 2 years ago

image Dear Dr.Luppino, When I am running Code_Aligned_Autoencoders.py, I find that after the final epoch, the 'ACC' obtained by final_evaluate() is much higher than that in the historical last epoch(such as the 'ACC' increases from 0.783 to 0.928 in this picture and I set 21 as the final epoch). Do you know the reason? It is strange. Thank you!

llu025 commented 2 years ago

The reason is simple: in the other epochs another function is used where a much simpler filter is applied to the difference image. This is because the dense CRF filtering method used in the final epoch is much more computationally expensive and therefore it would have slowed down the training.

I have to say, I did not try to change it and use the dense CRF method during all epochs. It might have a performance boost in terms of classification.