Reproduce “Animal Avatars: Reconstructing Animatable 3D Animals from Casual Videos” 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
Remy Sabathier, Niloy J. Mitra, David Novotny
Teams at Meta and University College London developed "Animal Avatars," a method that reconstructs high-quality, animatable 3D dog models from monocular consumer videos. The approach integrates dense surface supervision using Continuous Surface Embeddings with a parametric model, an implicit neural radiance field for detailed textures, and a robust optimization strategy including rigid motion factorization, leading to improved shape accuracy, temporal consistency, and appearance quality.
Key insights
- Dense surface supervision, specifically integrating Continuous Surface Embeddings with parametric models, provides a robust alternative to sparse keypoints for deformable object reconstruction, especially from challenging viewpoints.
- Combining the structural benefits of parametric models with the high-fidelity rendering capabilities of neural radiance fields can effectively capture both underlying shape and intricate appearance details.
- Decomposing camera motion from object motion significantly enhances the stability and quality of multi-parameter 3D reconstruction from video.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.