awesome-semantic-segmentation-pytorch

awesome-semantic-segmentation-pytorch copied to clipboard

I think the your ICNet implementation is not accurate. First, there is no PPM in the head. Second, no shared computation. In the implementation, forward twice lead to slower speed. https://github.com/hszhao/ICNet/issues/7

Also, the CFF module shoud fuse the output of previous one.

I re-implement the ICNet which is in my repo: https://github.com/lxtGH/Fast_Seg