Problem about the final result about sec-tensorflow HOT 3 CLOSED

I don't think there is any erros in your code. And since I update the evaluation code, all the miou of my experiment drops a little. And I just test my model using my new evaluation code, the final miou is also about 48%. So I just think there is a little error in my old evaluation code. Sorry to confuse you.
And for improving the the performance, I think the poly learning policy and using tag to clean all the probability of not-existing labels may do help. I just plan to do some more experiment to check my thought in a few days.

from sec-tensorflow.

One more thing, if you use fixed input_size for testing, you may get a better score.

from sec-tensorflow.

Thank you. I will also do some try. Waiting for your good news.

from sec-tensorflow.