Research map / 3D vision / Gaussian splatting
ArticulatedGS: Self-supervised Digital Twin Modeling of Articulated Objects using 3D Gaussian Splatting
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 ArticulatedGS, a self-supervised multi-step 3D Gaussian Splatting pipeline that reconstructs part-level appearance, part masks, and motion parameters of articulated objects from two static multi-view image sets.
Problem
Building digital twins of articulated objects (geometry, RGB appearance, part segmentation, and motion parameters) from RGB multi-view observations of two static articulation states, without 3D, motion, or semantic supervision; joint optimization of these parameters is ill-posed and unstable.
Approach
Decouples the joint optimization into stages: reconstruct the start-state 3D Gaussians, use a DeformNet (with ARAP regularization) to predict per-Gaussian displacements supervised by the end state, derive the movable mask and motion parameters with appearance and Chamfer geometry losses, then jointly refine Gaussians, mask, and motion parameters.
Key results
- On the PARIS dataset, Ours achieves lower motion angle error than Ditto and PARIS on most objects (e.g., 0.02 vs 0.73-89 for other methods on several synthetic objects) and far lower Chamfer distances than PARIS.
- Visual quality on the PARIS dataset: Ours reaches start-state PSNR around 36-43 dB on synthetic objects, above PARIS (33.8-38.3 dB) on most categories.
- The method is reported to take roughly one-third of the optimization time of depth-based methods (DTA, ArtGS) while achieving similar quality; the paper notes it is more stable than PARIS across trials.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ArticulatedGS instead of PARIS | ArticulatedGS · PARIS dataset, Foldchair | Ang Err: 7.90 -> 0.02 | Helped |
| use ArticulatedGS instead of PARIS | ArticulatedGS · PARIS dataset, Foldchair | CD-s: 9.12 -> 0.33 | Helped |
| use ArticulatedGS instead of PARIS | ArticulatedGS · PARIS dataset, Foldchair start state | PSNR: 37.95 -> 41.68 | Helped |
| add progressive optimization rather than jointly optimizing all parameters (ablation) | ArticulatedGS · PARIS dataset, Foldchair start state | PSNR: 31.49 -> 41.68 | Helped |
| add progressive optimization rather than jointly optimizing all parameters (ablation) | ArticulatedGS · PARIS dataset, Foldchair end state | LPIPS: 0.094 -> 0.059 | Helped |
| remove progressive optimization and jointly optimize all parameters (ablation) | ArticulatedGS · PARIS dataset, articulated-object part segmentation and motion estimation | part segmentation and motion estimation quality | Hurt |
Surprising findings
- The Vanilla single-stage joint optimization often collapses to labeling all Gaussians as one category, yielding nearly static learned motion.
Limitations
- Requires world coordinate alignment between the two states.
- Assumes corresponding areas across the two states have similar lighting.
- Real-object ground truth is hard to annotate, limiting quantitative evaluation; default setting assumes only one movable part.
Experimental setup
Datasets: PARIS dataset (synthetic and real objects) Shape2Motion (used by Ditto's pretrained model) self-collected articulated objects
Benchmarks: PARIS dataset self-constructed articulated object dataset
Metrics: PSNR SSIM LPIPS motion angle error motion position error geometric distance Chamfer distance (CD-s, CD-d, CD-w)
Baselines: Ditto PARIS Vanilla (single-stage joint optimization variant) Deformable 3D Gaussians (qualitative)
Builds on: 3D Gaussian Splatting PARIS Deformable 3D Gaussians as-rigid-as-possible deformation
Closest papers by meaning
- Building Interactable Replicas of Complex Articulated Objects via Gaussian Splatting (ICLR 2025 · similarity 0.93)
- REArtGS: Reconstructing and Generating Articulated Objects via 3D Gaussian Splatting with Geometric and Motion Constraints (NeurIPS 2025 · similarity 0.90)
- RigGS: Rigging of 3D Gaussians for Modeling Articulated Objects in Videos (CVPR 2025 · similarity 0.90)
- GauHuman: Articulated Gaussian Splatting from Monocular Human Videos (CVPR 2024 · similarity 0.89)
- TAGA: Self-supervised Learning for Template-free Animatable Gaussian Articulated Model (CVPR 2025 · similarity 0.88)
- SelfSplat: Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splatting (CVPR 2025 · 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.