Research map / 3D vision / Human pose estimation
MeshMamba: State Space Models for Articulated 3D Mesh Generation and Reconstruction
Yusuke Yoshiyasu, Leyuan Sun, Ryusuke Sagawa
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Mamba block instead of MLP (ablation) | MambaDiff3D · Unconditional human generation (SURREAL test) | 1-NNA (%): 73.7 -> 53.1 | Improved |
| use Mamba block instead of GNN (ablation) | MambaDiff3D · Unconditional human generation (SURREAL test) | 1-NNA (%): 74.2 -> 53.1 | Improved |
| use Mamba block instead of transformer (ablation) | MambaDiff3D · Unconditional human generation (SURREAL test) | 1-NNA (%): 53.6 -> 53.1 | No clear effect |
| serialize by part-IUV instead of SMPL connectivity order (ablation) | MambaDiff3D · Unconditional human generation (SURREAL test) | 1-NNA (%): 60.0 -> 54.4 | Improved |
| use seven part-IUV serializations instead of one (ablation) | MambaDiff3D · Unconditional human generation (SURREAL test) | 1-NNA (%): 54.4 -> 53.0 | Improved |
| combine SMPL and XYZ serialization instead of SMPL connectivity alone (ablation) | MambaDiff3D · Unconditional human generation (SURREAL test) | 1-NNA (%): 60.0 -> 53.1 | Improved |
| use six XYZ serializations instead of one alongside SMPL (ablation) | MambaDiff3D · Unconditional human generation (SURREAL test) | 1-NNA (%): 53.1 -> 53.5 | No clear effect |
| use random vertex ordering (ablation) | MeshMamba · Unconditional generation | training | Hurt |
| replace Mamba with transformer in diffusion layers (inference speed, 10475 vertices, A100, 250 DDIM steps) (ablation) | MambaDiff3D · Mesh generation speed | generation time: 28.1 sec -> 4.5 sec | Improved |
| replace Mamba with transformer (V100 generation time) (ablation) | MambaDiff3D · Mesh generation speed | generation time: 58.3 sec -> 6.6 sec | Improved |
| replace Mamba with transformer (training time, 6890 tokens) (ablation) | MeshMamba · Training speed | time per epoch: 100 minutes -> 18 min | Improved |
| fine-tune on UBody (ablation) | Mamba-HMR · UBody whole-body mesh recovery | PA-MVE All (mm): 26.3 -> 25.9 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- AiOS — AiOS: All-in-One-Stage Expressive Human Pose and Shape Estimation (CVPR 2024) · better on 2 of 2 shared results
- SMPLer-X-L — SMPLer: Taming Transformers for Monocular 3D Human Shape and Pose Estimation (TPAMI 2023) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- End-to-End Recovery of Human Shape and Pose (CVPR 2018 · similarity 0.89)
- SAM 3D Body: Robust Full-Body Human Mesh Recovery (CVPR 2026 · similarity 0.89)
- Recovering 3D Human Mesh From Monocular Images: A Survey (TPAMI 2023 · similarity 0.88)
- SCULPT: Shape-Conditioned Unpaired Learning of Pose-dependent Clothed and Textured Human Meshes (CVPR 2024 · similarity 0.88)
- GenHMR: Generative Human Mesh Recovery (AAAI 2025 · similarity 0.87)
- AiOS: All-in-One-Stage Expressive Human Pose and Shape Estimation (CVPR 2024 · similarity 0.86)
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.
