NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:1710
Title:Multi-mapping Image-to-Image Translation via Learning Disentanglement

While the reviewer scores diverged a bit, the reviews were actually in good agreement. The reviewers found the paper to be novel and significant because it proposes a novel unified framework multi-domain and multi-modality image to image (I2I) translation (while previous works focus only one of them). However the technical novelty is limited in the sense that this is accomplished by combining losses and disentanglement approaches that have been proposed in the previous works. There were also concerns that the experiments are performed on two domains only, and it is not clear that the method will work between domains where the difference is not only in style. There were also various technical issues that the authors addressed in the rebuttal. After discussion, the reviewers split, with one positive, one marginally positive, and one marginally negative. This made the paper a very borderline case. After discussion between ACs it was decided that the novelty and importance of the problem as well as the soundness of the method justify publication, even though the technical novelty is limited.