Research map / 3D vision / Gaussian splatting
RigGS: Rigging of 3D Gaussians for Modeling Articulated Objects in Videos
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
RigGS rigs 3D Gaussians for monocular videos of articulated objects by building a sparse 3D skeleton from skeleton-aware deformation nodes, then driving the Gaussians with learnable skinning and pose-dependent detail deformation, without template priors.
Problem
Modeling articulated objects from monocular 2D videos so they support novel view synthesis plus editing, driving, and reposing, without template skeleton priors or predefined skeleton trees.
Approach
Use a skeleton-aware node-controlled deformation field over canonical 3D Gaussians to produce candidate skeleton nodes, simplify them via a heuristic (FPS, minimum spanning tree, geometric/motion/DINOv2 semantic cues) into a sparse skeleton, then bind Gaussians with learnable LBS skinning weights and a pose-dependent detail deformation.
Key results
- On D-NeRF, RigGS reaches 40.82 PSNR / 0.996 SSIM / 0.0112 LPIPS, slightly below SC-GS (43.04) but well above 4D-GS (33.25) and AP-NeRF (30.94).
- On DG-Mesh, RigGS scores 37.65 PSNR versus 38.96 for SC-GS and 33.96 for 4D-GS.
- On ZJU-MoCap, RigGS reaches 33.54 PSNR versus 25.62 for AP-NeRF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RigGS instead of SC-GS | Dynamic object novel-view synthesis · D-NeRF dataset | PSNR: 43.04 -> 40.82 | Hurt |
| Use RigGS instead of SC-GS | Dynamic object novel-view synthesis · D-NeRF dataset | LPIPS: 0.0066 -> 0.0112 | Hurt |
| Use RigGS instead of SC-GS | Dynamic object novel-view synthesis · DG-Mesh dataset | PSNR: 38.96 -> 37.65 | Hurt |
| Use RigGS instead of SC-GS | Dynamic object novel-view synthesis · DG-Mesh dataset | LPIPS: 0.0136 -> 0.0169 | Hurt |
| Use RigGS instead of AP-NeRF | Dynamic object novel-view synthesis · ZJU-MoCap | PSNR: 25.62 -> 33.54 | Helped |
| Use RigGS instead of AP-NeRF | Dynamic object novel-view synthesis · ZJU-MoCap | LPIPS: 0.0934 -> 0.0327 | Helped |
| Use isotropic instead of anisotropic 3D Gaussians (ablation) | RigGS · D-NeRF dataset | Rendering quality on new poses | Helped |
| Remove the 2D projection loss (ablation) | RigGS skeleton-driven dynamic model · D-NeRF dataset | Skeleton alignment | Hurt |
| Use a fixed rather than adaptive 2D projection-loss weight (ablation) | RigGS skeleton-driven dynamic model · D-NeRF dataset | Skeleton alignment | Hurt |
Limitations
- Effectiveness limited with sparse viewpoints, inaccurate camera poses, or excessive motion.
- Pose-related appearance is not modeled.
- Rendering accuracy slightly below SC-GS on the benchmarks.
Experimental setup
Datasets: D-NeRF DG-Mesh ZJU-MoCap
Benchmarks: D-NeRF DG-Mesh ZJU-MoCap
Metrics: PSNR SSIM LPIPS
Baselines: D-NeRF TiNeuVox 4D-GS SC-GS AP-NeRF
Builds on: 3D Gaussian Splatting SC-GS AP-NeRF LBS
Closest papers by meaning
- ArticulatedGS: Self-supervised Digital Twin Modeling of Articulated Objects using 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- Template-free Articulated Gaussian Splatting for Real-time Reposable Dynamic View Synthesis (NeurIPS 2024 · similarity 0.89)
- Learning Implicit Representation for Reconstructing Articulated Objects (ICLR 2024 · similarity 0.89)
- REACTO: Reconstructing Articulated Objects from a Single Video (CVPR 2024 · similarity 0.89)
- Building Interactable Replicas of Complex Articulated Objects via Gaussian Splatting (ICLR 2025 · similarity 0.89)
- REArtGS: Reconstructing and Generating Articulated Objects via 3D Gaussian Splatting with Geometric and Motion Constraints (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.