Open Sur-4 opened 5 days ago
Hi @Sur-4 ,
I just found that the ground truth (targets folder) of post-disaster images is building damage classification, not non-building and building class as we have for pre-disaster images. So, we might not be able to do localization for post-disaster images.
Hi japan, I have saved the output from the pre-disaster localisation using U-net https://drive.google.com/drive/folders/1vRqcXKmVdeudErPdEiXvF02V9_w_mSI1?usp=sharing this needs to be as input in the next u-net classification model along with pre-disaster images, post-disaster images and building classes mask but we'll have to figure out the inconsistencies as you mentioned first. anyway please let me know if you manage to debug that, i will do the same :D
Hi japan, there seems to be an issue with label encoding of 'post'. the prob is, when i try to visualize, the classes should be vice-versa right, but 1 seems to be blank. could you pls try and check if it the same for you as well? i tried my best to debug this but couldn't spot it. P.S 'pre' condition works well, so i suspect some bug in the 'post' condition/mask_polygon