Reproduce “Self-Learning Expression Deformations for Data-Efficient Gaussian Avatars” on OpenResearch

About this paper

Jiahao Yang, Xiaohang Yang, Qing Wang, Yilan Dong, Gregory Slabaugh, Shanxin Yuan

SAGE (Self-Adaptive Gaussian Expression), a framework from Queen Mary University of London, creates high-fidelity, animatable Gaussian avatars using minimal input data by employing self-learning expression deformations. It achieves superior performance in self-reenactment tasks across various data-limited settings, outperforming baselines that require extensive multi-frame sequences.

Key insights

• Jointly optimizing an SDF with 2DGS surfels, complemented by parametrized Gaussian attributes, ensures surface-aligned and geometry-aware distributions, improving spatial and transformation consistency.
• Self-supervised learning through geometric and appearance consistency constraints can effectively replace the need for extensive, supervised training sequences for dynamic expression deformation.
• A Gaussian surfel correction mechanism, which adjusts scales and rotations based on estimated surface stretching and principal curvatures, helps maintain geometric integrity during dynamic deformation.

Summary by alphaXiv.