GaussiAnimate: Reconstruct and Rig Animatable Categories with Level of Dynamics

How to read this

Category

research paper, automatic rigging and reanimation of Gaussian reconstructed animatable objects

Contributions

• Introduces Skelebones, a scaffold skin rigging system that compresses temporally consistent deformable Gaussian reconstructions into free form bones
• Extracts a Mean Curvature Skeleton from the canonical Gaussian representation and refines it temporally, giving the free form bones a proper kinematic structure
• Binds skeleton and bones through a non parametric partwise motion matching algorithm, PartMM, rather than a learned or hand authored skinning step
• Reports quantitative reanimation gains over linear blend skinning and bag of bones baselines on unseen poses, and shows PartMM generalizes to both Gaussian and mesh representations even in a low data regime

Context

This paper sits in the recent line of work reconstructing animatable categories, animals and generic objects, directly from casual video via free form bones and Gaussian splats, and it addresses a gap those methods share: free form bones capture non rigid deformation well but lack the kinematic structure needed for intuitive rigger style control. No builds_on entries are listed in the archive, but the paper positions itself directly against prior free form bone and bag of bones reconstruction methods it benchmarks against.

Correctness

No PDF for this entry was available in the archive, so this guide is grounded in secondhand web summaries of the arXiv abstract rather than a firsthand read of the paper. Those summaries report 17.3 percent PSNR gains over linear blend skinning, 21.7 percent over Bag of Bones, and a 48.4 percent RMSE improvement over robust linear blend skinning in a roughly 1000 frame low data regime, but these specific numbers should be treated as reported rather than independently confirmed until the tables themselves are read.

Clarity

Not directly assessed since the PDF was unavailable, but the reported abstract summary reads as a fairly standard modern 3D vision paper structure, a method section plus quantitative ablations, aimed at researchers already familiar with Gaussian splatting and free form bone rigging literature.

How to read it

Five minute pass: read the arXiv abstract at arxiv.org/abs/2604.08547 to confirm the Skelebones pipeline, Bones, Skeleton, Binding, and the headline numbers. Second pass: get the PDF and read the method section on Mean Curvature Skeleton extraction and PartMM binding, since that is the actual novel contribution rather than the Gaussian compression step. Third pass is worth it only for someone building an automatic rigging pipeline off reconstructed Gaussians, who should also chase down the linear blend skinning and Bag of Bones baselines cited for comparison to judge whether the reported gains hold on their own data.