Implementing this paper < https://arxiv.org/abs/1903.03129 > for a computer vision application.