Learning Riemannian Metrics for Interpolating Animations

Sarah Kushner, Vismay Modi, Nina Miolane

Keywords: lie groups, animation, rotations, interpolation, motion representation, compression

Abstract: We propose a new family of geodesic interpolation techniques to perform upsampling of low frame rate animations to high frame rates. This approach has important applications for: (i) creative design, as it provides a diversity of interpolation methods for digital animators; and (ii) compression, as an original high frame rate animation can be recovered with high accuracy from its subsampled version. Specifically, we upsample low frame rate animations by interpolating the rotations of an animated character's bones along geodesics in the Lie group $SO(3)$ for different invariant Riemannian metrics. For compression, we propose an optimization technique that selects the Riemannian metric whose geodesic most faithfully represent the original animation. We demonstrate the advantages of our approach compared to existing interpolation techniques in digital animation.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2933