Research map / 3D vision / Face and shape
REACTO: Reconstructing Articulated Objects from a Single Video
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.
Reconstructs piece-wise rigid articulated objects from a single casual monocular video using a canonical NeRF and a Quasi-Rigid Blend Skinning deformation model with bone-based rigging, quasi-sparse skinning weights, and geodesic point assignment.
Problem
Reconstructing general articulated 3D objects (piece-wise rigid, with adjacent movable parts) from a single monocular casual video, where existing blend-skinning deformation models designed for humans/animals produce seam artifacts, tearing, and motion errors.
Approach
Define the rig on bones (part centroids) instead of joints, make Gaussian skinning weights quasi-sparse via a temperature factor, and use geodesic distance on the canonical mesh for point-to-bone/joint assignment with a sparse skinning loss, combining rigid skinning's rigidity with dual quaternion blend skinning's joint flexibility.
Key results
- On synthetic USB, stapler, and scissors, REACTO achieves CD 15.3/14.3/14.0 and F(10%) 78.6/75.5/78.2, outperforming BANMo, MoDA, and PPR on all metrics.
- Ablation: REACTO beats displacement field (NASAM), Real-NVP (CaDeX), and binarized rigid skinning on all synthetic objects, e.g., USB CD 15.3 vs 19.7, 17.6, 16.3.
- Qualitative results on real-world videos show fewer seam, tearing, and over-smoothing artifacts than baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use REACTO instead of the strongest listed baseline | 3D articulated object reconstruction · USB | CD: 17.1 -> 15.3 | Helped |
| Use REACTO instead of the strongest listed baseline | 3D articulated object reconstruction · stapler | CD: 16.8 -> 14.3 | Helped |
| Use REACTO instead of the strongest listed baseline | 3D articulated object reconstruction · scissors | CD: 14.8 -> 14.0 | Helped |
| Use a displacement field instead of QRBS (ablation) | REACTO · USB | F(10%): 59.5 -> 78.6 | Hurt |
| Use Real-NVP instead of QRBS (ablation) | REACTO · USB | F(10%): 70.7 -> 78.6 | Hurt |
| Use rigid skinning instead of QRBS (ablation) | REACTO · USB | F(10%): 73.5 -> 78.6 | Hurt |
| Use a displacement field instead of QRBS (ablation) | REACTO · stapler | F(10%): 63.8 -> 75.5 | Hurt |
| Use Real-NVP instead of QRBS (ablation) | REACTO · stapler | F(10%): 70.4 -> 75.5 | Hurt |
| Use rigid skinning instead of QRBS (ablation) | REACTO · stapler | F(10%): 72.8 -> 75.5 | Hurt |
| Use a displacement field instead of QRBS (ablation) | REACTO · scissors | F(10%): 58.6 -> 78.2 | Hurt |
| Use Real-NVP instead of QRBS (ablation) | REACTO · scissors | F(10%): 63.6 -> 78.2 | Hurt |
| Use rigid skinning instead of QRBS (ablation) | REACTO · scissors | F(10%): 76.4 -> 78.2 | Hurt |
Surprising findings
- Binarized rigid skinning performs comparably to the full method on synthetic data, though it produces seam artifacts on real videos.
Limitations
- Casual videos give only partial views, so the unseen side of the surface may be poorly reconstructed.
- Relies on off-the-shelf mask, optical flow, and feature predictors plus sparse manual camera pose annotations (about 4 per video).
Experimental setup
Datasets: PartNet-Mobility (synthetic renders: USB, stapler, scissors) real-world phone videos of laptops, staplers, scissors, faucets, nail clippers, glasses
Benchmarks: PartNet-Mobility synthetic renders (USB, stapler, scissors)
Metrics: Chamfer distance F-score at 10% F-score at 5%
Baselines: BANMo MoDA PPR NASAM displacement field Real-NVP (CaDeX) Rigid skinning
Builds on: BANMo PPR MoDA NASAM Lab4D
Closest papers by meaning
- Learning Implicit Representation for Reconstructing Articulated Objects (ICLR 2024 · similarity 0.91)
- REArtGS: Reconstructing and Generating Articulated Objects via 3D Gaussian Splatting with Geometric and Motion Constraints (NeurIPS 2025 · similarity 0.89)
- RigGS: Rigging of 3D Gaussians for Modeling Articulated Objects in Videos (CVPR 2025 · similarity 0.89)
- Reconstructing Humans with a Biomechanically Accurate Skeleton (CVPR 2025 · similarity 0.88)
- S3O: A Dual-Phase Approach for Reconstructing Dynamic Shape and Skeleton of Articulated Objects from Single Monocular Video (ICML 2024 · similarity 0.88)
- SINGAPO: Single Image Controlled Generation of Articulated Parts in Objects (ICLR 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.