Research map / 3D vision / Gaussian splatting
Template-free Articulated Gaussian Splatting for Real-time Reposable Dynamic View Synthesis
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.
Proposes a template-free method that reconstructs dynamic articulated objects with 3D Gaussians and superpoints, discovers a skeleton from superpoint motion, and refines it with a kinematic model for real-time reposable novel view synthesis.
Problem
Capturing skeleton models of articulated objects from multi-view videos without object-specific templates or pose annotations, while keeping high-fidelity novel view synthesis and real-time rendering for reposing.
Approach
Use 3D Gaussians bound to superpoints (treated as rigid parts) with LBS deformation; discover joints and a minimum-spanning-tree skeleton from superpoint relative motion, then optimize a kinematic tree model; adaptive prune/densify/merge of superpoints with ARAP and smoothness/sparsity regularizers keeps the superpoint count small.
Key results
- On D-NeRF, the method reports PSNR 38.80 at 800x800 with 103.98 FPS and 39.23 PSNR at 400x400 with 110.90 FPS, versus AP-NeRF's 30.91 PSNR at 1.33 FPS.
- On Robots, PSNR 34.34 and 137.76 FPS at 512x512, outperforming AP-NeRF (32.45 PSNR, 0.89 FPS).
- On ZJU-MoCap, achieves 198.23 FPS but lower PSNR (29.11) than WIM (31.08) and AP-NeRF (29.60).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed articulated Gaussian method instead of AP-NeRF | Articulated novel-view synthesis · Robots | PSNR: 32.45 -> 34.34 | Helped |
| Use the proposed articulated Gaussian method instead of AP-NeRF | Articulated novel-view synthesis · Robots | FPS: 0.89 -> 137.76 | Helped |
| Use the proposed articulated Gaussian method instead of WIM | Articulated novel-view synthesis · ZJU-MoCap | PSNR: 31.08 -> 29.11 | Hurt |
| Use the proposed articulated Gaussian method instead of AP-NeRF | Articulated novel-view synthesis · ZJU-MoCap | FPS: 1.31 -> 198.23 | Helped |
| Vary the initial number of superpoints (ablation) | Ours · D-NeRF (hellwarrior) | PSNR: 39.61, 39.70, 39.83, 39.69, 39.89, 39.56, 39.72, 40.43 | No clear effect |
Surprising findings
- On ZJU-MoCap real-world data, the method has lower PSNR than WIM and AP-NeRF, attributed to imperfect camera calibration.
Limitations
- Learned skeleton is restricted to the kinematic motion space in the input video, so extrapolation to unseen poses may be erroneous.
- Inaccurate camera poses, large motion, or long sequences cause reconstruction failures.
- Handles only a single articulated object; multi-component objects or scenes are not addressed.
Experimental setup
Datasets: D-NeRF Robots ZJU-MoCap
Benchmarks: D-NeRF Robots ZJU-MoCap
Metrics: PSNR SSIM LPIPS FPS
Baselines: D-NeRF TiNeuVox HexPlane K-Planes Deformable 3D-GS 4D-GS SP-GS SC-GS WIM AP-NeRF
Builds on: 3D Gaussian Splatting SC-GS Watch-It-Move AP-NeRF linear blend skinning
Closest papers by meaning
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.90)
- SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes (CVPR 2024 · similarity 0.90)
- RigGS: Rigging of 3D Gaussians for Modeling Articulated Objects in Videos (CVPR 2025 · similarity 0.89)
- Superpoint Gaussian Splatting for Real-Time High-Fidelity Dynamic Scene Reconstruction (ICML 2024 · similarity 0.89)
- Swift4D: Adaptive divide-and-conquer Gaussian Splatting for compact and efficient reconstruction of dynamic scene (ICLR 2025 · similarity 0.88)
- Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction (CVPR 2024 · similarity 0.88)
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.