NeurIPS 2020

An Analysis of SVD for Deep Rotation Estimation


Meta Review

All reviewers concur that it is surprising that SVD as presented in this paper is not more widely used in DL applications. The primary value of the paper is really the empirical evidence that SVD is valuable. The reviewers disagree somewhat on the level of surprise in the derived results, but it remains valuable to include them in the paper, as a reference for readers who wish to make use of the empirical findings. The abstract's claim "we provide a theoretical analysis that shows SVD is the natural choice..." is rather empty. I might just as well show quaternions are "natural", by deriving some other quantities. I think it would be much better to write this as "We summarise some theoretical properties of SVD as used for representation of (and projection onto) the rotation group."