We currently use the author's original implementation, which add another (unnecessary) dependency to cv2. It would be much nicer to have this as a differentiable noise model purely implemented in pytorch, merging it with the semantic segmentation pipeline.
We currently use the author's original implementation, which add another (unnecessary) dependency to cv2. It would be much nicer to have this as a differentiable noise model purely implemented in pytorch, merging it with the semantic segmentation pipeline.