Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
The paper proposes a new graph neural network architecture for vision tasks such as relighting and face swapping. By incorporating filtering and propagation mechanisms from classic graph Laplacian techniques, they recreate this behavior in deep neural networks. Through qualitative experiments, the paper argues that this leads to improved performance on the vision tasks considered. The reviewers generally agree that this is an interesting, novel approach, but have questions regarding evaluation. Since the results are qualitative, it is difficult to gain insight into general differences between the new method and previous ones. The reviewers encourage the additional results, such as the user study, presented in the authors' response to be added to the manuscript.