NeurIPS 2020

CaSPR: Learning Canonical Spatiotemporal Point Cloud Representations

Meta Review

Reviewers all agreed that this is a strong submission. The proposed problem is novel and interesting. Reviewers called the method “elegant and well designed.” The paper is very well written. Reviewers also appreciated that “the experiments section does a great job at highlighting the effects of individual design choices.” The paper demonstrates the applicability of the proposed method on a number of downstream tasks. The main limitations of the paper mentioned by reviewers are: the paper builds on many methods from previous work and has limited technical novelty the method requires strong supervision that is difficult to obtain in the real world the method operates at the object level and cannot currently handle an entire scene a separate model is trained per object category Overall the reviewers all agreed that this is a very strong submission.