BLSeg
BLSeg copied to clipboard
Understanding fcn.py
Thanks for creating the repo. Can you please explain lines 60-64 of fcn.py?
score_pool4 = score_pool4[:, :, 5:5 + score2.size(2), 5:5 + score2.size(3)]
score16 = self.score16(score2 + score_pool4)
x = score16[:, :, 27:27 + x.size(2), 27:27 + x.size(3)]
It means crop
The feature map of score_pool4 is larger than score2's out
Here is the architecture of FCN https://ethereon.github.io/netscope/#/preset/fcn-8s-pascal
The 'crop-pool4' is doing the same things
I hope it helps you