IMELE
IMELE copied to clipboard
Post processing
Thanks for your work. Did you perform post-processing to merge the predicted height image patches and generate the seamless height map?
That would be a good idea Rain. I always find the inconsistent DSM patch from one to another although I train the model with 250 pixel overlap throughout the whole tile. I think it is because the problem arises from single depth(height) estimation ambiguous.
I did not merge them in a seamless fashion so the large error always comes in the edge. You are welcome to improve it.