Reproduce “Motion Inbetweening via Deep $Δ$-Interpolator” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Boris N. Oreshkin, Antonios Valkanas, Félix G. Harvey, Louis-Simon Ménard, Florent Bocquelet, Mark J. Coates
Researchers from Unity Technologies and McGill University developed the Deep ∆-Interpolator, a Transformer-based model for 3D character motion in-betweening. This method predicts motion as a residual relative to a localized baseline, achieving state-of-the-art accuracy and enhanced robustness against out-of-distribution data on the LaFAN1 and Anidance datasets.
Key insights
- Operating in a 'delta-mode' by predicting residuals relative to a baseline significantly improves the accuracy and effectiveness of deep learning interpolators for motion synthesis.
- Utilizing a local reference frame for both network inputs and outputs enhances model robustness against out-of-distribution data and domain shifts, a common issue with global coordinate systems.
- The delta-regime remains effective even with a simplified baseline (like the last known frame), demonstrating flexibility and reducing computational dependency on complex initial interpolations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.