NeurIPS 2020

SDF-SRN: Learning Signed Distance 3D Object Reconstruction from Static Images


Meta Review

Two knowledgeable referees supported accept and two knowledgeable referees supported reject. Lack of clarity + emphasis in the writing of the paper had created doubts in reviewers' minds about how and why the method performs as well as it does. The reviewers also had questions about the performance of baselines (DVR in particular). R2 has misunderstood the paper in at least 2 places and despite the author's rebuttal pointing them to their misunderstandings, they have failed to adjust their judgement. I therefore disregard their review. I also believe some of the other reviewers' concerns about the method's surprisingly good performance have been and can be addressed by the authors in revisions. Given the other reviewer comments, and my own reading of he paper, I believe the paper should be accepted as a poster. The paper addresses a significant problem, proposes a non-trivial novel solution, is well written and it is SOTA. I highly encourage the authors to address these issues to the fullest extent possible before publication.