Research map / 3D vision / 3D generation
Skinned Motion Retargeting with Dense Geometric Interaction Perception
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 MeshRet, a single-stage skinned motion retargeting framework that models dense geometric interactions via semantically consistent sensors and a dense mesh interaction field, trained on Mixamo and a new ScanRet dataset.
Problem
Skinned motion retargeting often produces interpenetration, jitter and contact mismatches because skeleton-level retargeting and geometry correction conflict; existing methods ignore body geometry or correct it in a separate stage.
Approach
Establish dense mesh correspondence across different topologies using semantically consistent sensors derived MAIT-style from skeleton bones, encode pairwise sensor interactions as a spatio-temporal DMI field, and align the source and target DMI fields during a single-pass transformer retargeting network.
Key results
- MeshRet reduces MSE by 65% vs PMnet, 29% vs R2ET and 8% vs SAN on ScanRet.
- MeshRet achieves contact error 0.772 on Mixamo+ and 0.284 on ScanRet, the lowest among compared methods.
- In a 600-comparison user study, about 81% of preferences favored MeshRet over baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a mixture of proximal and distal sensor pairs instead of only proximal pairs (ablation) | MeshRet · Mixamo+ | Contact Error: 0.800 -> 0.772 | Helped |
| use a mixture of proximal and distal sensor pairs instead of only proximal pairs (ablation) | MeshRet · Mixamo+ | Penetration (%): 3.35 -> 3.45 | Hurt |
| add proximal sensor pairs instead of using only farthest pairs (ablation) | MeshRet · Mixamo+ | Contact Error: 1.642 -> 0.772 | Helped |
| add proximal sensor pairs instead of using only farthest pairs (ablation) | MeshRet · Mixamo+ | Penetration (%): 5.37 -> 3.45 | Helped |
| replace the DMI consistency loss with distance matrix loss (ablation) | MeshRet · Mixamo+ | Contact Error: 0.772 -> 2.568 | Hurt |
| replace the DMI consistency loss with distance matrix loss (ablation) | MeshRet · Mixamo+ | Penetration (%): 3.45 -> 4.69 | Hurt |
| use half the sample points in the SCS φ space (ablation) | MeshRet · Mixamo+ | Penetration (%): 3.45 -> 3.90 | Hurt |
| use half the sample points in the SCS l space (ablation) | MeshRet · Mixamo+ | Contact Error: 0.772 -> 0.805 | Hurt |
| use 75% proximal sensor pairs instead of the 50% mixture (ablation) | MeshRet · Mixamo+ | Contact Error: 0.772 -> 0.909 | Hurt |
| use 25% proximal sensor pairs instead of the 50% mixture (ablation) | MeshRet · Mixamo+ | Contact Error: 0.772 -> 0.781 | No clear effect |
Surprising findings
- A distance matrix loss from prior work provided no meaningful supervisory signal, as distance is non-directional.
- Using only the farthest sensor pairs significantly degrades contact preservation and increases interpenetration.
Limitations
- Depends on clean-contact inputs; severely interpenetrating source clips yield poor results.
- SCS extraction can be compromised by noisy meshes, such as complex clothing.
- Cannot handle characters with missing limbs.
Experimental setup
Datasets: Mixamo ScanRet
Benchmarks: ScanRet Mixamo+
Metrics: MSE local MSE contact error penetration ratio human preference
Baselines: Copy PMnet SAN R2ET Zhang et al. 2023 semantics
Builds on: PMnet MAIT PointNet FaceFormer alignment mask Zhang et al. 2023 skinned motion retargeting (R2ET)
Closest papers by meaning
- Semantics-aware Motion Retargeting with Vision-Language Models (CVPR 2024 · similarity 0.84)
- InterAct: Advancing Large-Scale Versatile 3D Human-Object Interaction Generation (CVPR 2025 · similarity 0.83)
- Reconstructing Close Human Interaction with Appearance and Proxemics Reasoning (CVPR 2025 · similarity 0.83)
- ContactField: Implicit Field Representation for Multi-Person Interaction Geometry (NeurIPS 2024 · similarity 0.83)
- AvatarGO: Zero-shot 4D Human-Object Interaction Generation and Animation (ICLR 2025 · similarity 0.83)
- DenseMatcher: Learning 3D Semantic Correspondence for Category-Level Manipulation from a Single Demo (ICLR 2025 · similarity 0.83)
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.