I am currently trying to train a UNet model on the dronedeploy segmentation dataset. My problem is that I only get outputs like this from my model:
This means that in the output maps I only get 0s for 1st, 2nd, 4th and 6th class and only 1s for 3rd and 5th class. I've already taken a closer look at the outputs and noticed that before applying my threshold of 0.5, but after applying softmax in a map, I have the same numbers everywhere. The average DICE seems to be pretty good at 0.71 at max.
I am using pytorch and cross entropy loss. I already tried other class weights and DICE loss but got similar results.