ebf99bb5df6533b6dd9180a59034698d-AuthorFeedback.pdf

R1: Pose prior makes little difference in Tab. 2. Yes, the benefit of the n.f. This is achieved automatically, by the best-of-M loss: 'guessing' only a single hypothesis would be20 disadvantageous compared to outputting M diverse guesses as that has a higher chance that one would have low21 error. Still, previous best-of-M implementations fail to reliably converge to this optimal solution, resulting in some22 modes to degenerate. An advantage of our loss (Eq 5) is that it helps to keep all modes active (L39-53). R2:23 How is the n.f.