I wanted to implement PointRend Net from scratch (https://arxiv.org/pdf/1912.08193v2.pdf) to predict a binary mask. Can anyone help me understand how to get the coarse grained and the fine grained features mentioned in the paper?
Any suggestions about the implementation are highly appreciated.
For the coarse grained features my understanding from reading the paper is that they would be the output of MaskRCNN Mask head at a lower resolution. How would you obtain the fine grained features?