NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:3907
Title:PointDAN: A Multi-Scale 3D Domain Adaption Network for Point Cloud Representation


		
The submission received mixed ratings prior and after discussion phase. The final scores are 5,6,7 with a tendency of lowering the 7 to a 6. The reviewers acknowledge the quality of the paper, the new dataset and that the technique is a good approach to domain adaption in 3D. The paper is among the first ones, the statements about being the first should be revised given the related literature pointed out by R1. R2 has doubts whether the method is too specificly tuned to PointNet while R3 believes PointNet is an already well enough established method that even if that would be the case the method would still be of interest. The empirical results are varying quite a bit over the classes, the authors answer to this point with a new experiment. This was acknowledged by the reviewers in the discussion and should be added to the final paper. Overall the positive points outweight the doubts about whether the feature alignment is a necessary step to perform 3D domain adaption. The paper presents a valid contribution, with an attempt of a theoretical justification and a novel datasets. Empirical results are sufficient to back the claims made by the paper.