Partial Convolution based Padding

Deep Learning
1

Has someone tried the idea from this paper?
https://arxiv.org/abs/1811.11718

Intuitively it makes sense, but it is one of the cases where it almost sounds too good to be true.