AnimalGS: 4D Animal Reconstruction from Monocular video with 3D Gaussian Splatting

ICLR 2026 Conference Submission19404 Authors

Keywords: Animal; 3D Gaussian Splatting; 4D Reconstruction

TL;DR: A 4D reconstruction method from monocular video for animals

Abstract:

Reconstructing 4D animals from monocular videos is challenging due to large inter-species variation, complex articulations, and the lack of reliable templates.
We introduce AnimalGS, a test-time optimization framework built on a 3D Gaussian Splatting representation for high-fidelity 4D reconstructions from single videos.
Grounded in the insight that robust reconstruction emerges from pose-guided optimization rather than strict shape priors, AnimalGS treats priors as coarse initializations and integrates joint-aware and symmetry-aware designs to progressively disentangle motion and appearance. This leads to empirically strong generalization across diverse species and robustness to mismatching with shape priors.
Extensive experiments demonstrate the superior performance of our approach in geometry, motion, and temporal consistency across a wide variety of animal species.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 19404