InstantGeoAvatar: Effective Geometry and Appearance Modeling of Animatable Avatars from Monocular Video

TL;DR

InstantGeoAvatar is a fast and stable method for creating detailed, animatable 3D human avatars from monocular videos, using a novel regularization scheme for SDFs that improves quality and reduces training time.

Contribution

It introduces a geometry-aware SDF regularization that stabilizes hash grid optimization, enabling rapid and high-quality avatar reconstruction from monocular videos.

Findings

Outperforms previous methods in geometry accuracy and view synthesis.

Achieves training in as little as five minutes, significantly faster than prior approaches.

Provides stable and detailed avatar models suitable for interactive applications.

Abstract

We present InstantGeoAvatar, a method for efficient and effective learning from monocular video of detailed 3D geometry and appearance of animatable implicit human avatars. Our key observation is that the optimization of a hash grid encoding to represent a signed distance function (SDF) of the human subject is fraught with instabilities and bad local minima. We thus propose a principled geometry-aware SDF regularization scheme that seamlessly fits into the volume rendering pipeline and adds negligible computational overhead. Our regularization scheme significantly outperforms previous approaches for training SDFs on hash grids. We obtain competitive results in geometry reconstruction and novel view synthesis in as little as five minutes of training time, a significant reduction from the several hours required by previous work. InstantGeoAvatar represents a significant leap forward towards achieving interactive reconstruction of virtual avatars.