Can you achieve the mIou and speed of the paper?

First of all, thank you for sharing your code. I have a question: how does shared convolution update parameters? Is it based on gradient flow?

Would you mind providing your backbone pretrained files? For limited storage and GPU processing power in our lab, we want to know whether it is convient for you to provide us with pretrained model of backbone network? It will be a lot of help, Thanks a lot to your excellent job in coding!