Research map / 3D vision / Human pose estimation
Hamba: Single-view 3D Hand Reconstruction with Graph-guided Bi-Scanning Mamba
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.
Hamba proposes a graph-guided Mamba (bidirectional state-space scanning over a few sampled joint tokens) for single-view 3D hand mesh reconstruction, reaching SOTA on FreiHAND, HO3Dv2/v3 and HInt.
Problem
Single-RGB 3D hand mesh reconstruction is hard under articulation, self-occlusion, truncation and hand-object interaction; attention-based transformers model joint spatial relations inefficiently and use many tokens.
Approach
Reformulate Mamba's scanning as graph-guided bidirectional scanning over ~21 joint-sampled tokens (plus a global mean token), combining GCN layers with SS2D Mamba blocks in a GSS block, and fuse GSS, sampled, 2D-joint and global features before MANO regression.
Key results
- On FreiHAND with TTA, Hamba reaches PA-MPVPE 5.3 mm and F@15mm 0.992, versus HaMeR at 5.7 mm and 0.990.
- On HO3Dv2 Hamba reports PA-MPVPE 7.7 mm and AUC_J 0.850, the best among listed methods (HaMeR-170k 7.9 and 0.848).
- Hamba uses 22 tokens versus 192 for attention-based decoders, an 88.5% token reduction, with 46.1% lower decoder runtime.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Hamba instead of HaMeR | 3D hand reconstruction · FreiHAND | PA-MPVPE: 5.7 -> 5.5 | Helped |
| use Hamba instead of HaMeR-170k | 3D hand reconstruction · HO3Dv2 | PA-MPJPE: 7.6 -> 7.5 | Helped |
| use Hamba instead of AMVUR | 3D hand reconstruction · HO3Dv3 | PA-MPJPE: 8.7 -> 6.9 | Helped |
| add Token Sampler Branch to fusion (ablation) | Hamba · FreiHAND | F@5mm: 0.722 -> 0.738 | Helped |
| add 2D Joints Feature Branch to fusion (ablation) | Hamba · FreiHAND | F@5mm: 0.718 -> 0.738 | Helped |
| add GSS Token Branch to fusion (ablation) | Hamba · FreiHAND | F@5mm: 0.717 -> 0.738 | Helped |
| add Global Mean Token Branch to fusion (ablation) | Hamba · FreiHAND | F@5mm: 0.680 -> 0.738 | Helped |
| add Token Sampler (ablation) | Hamba · FreiHAND | F@5mm: 0.717 -> 0.738 | Helped |
| use bidirectional instead of unidirectional scanning (ablation) | Hamba · FreiHAND | F@5mm: 0.718 -> 0.738 | Helped |
| add GCN (ablation) | Hamba GSS block · FreiHAND | F@5mm: 0.673 -> 0.738 | Helped |
| use graph-guided joint scan order instead of shuffled joint order (ablation) | Hamba · FreiHAND | F@5mm: 0.680 -> 0.738 | Helped |
| add Mamba blocks (SS2D, Layer Norm, and FFN) (ablation) | Hamba GSS block · FreiHAND | F@5mm: 0.675 -> 0.738 | Helped |
Surprising findings
- Replacing the SS2D blocks with attention inside the same GCN decoder gives a smaller gain, but removing Mamba blocks while keeping GCN gives an equal drop to removing GCN, which the authors read as both being essential.
Limitations
- Does not model temporal features in videos.
- May not cover all in-the-wild situations; failures on palm orientation and motion blur.
Experimental setup
Datasets: FreiHAND HO3D MTC RHD InterHand2.6M H2O3D DexYCB COCO-Wholebody Halpe MPII NZSL HInt
Benchmarks: FreiHAND HO3Dv2 HO3Dv3 HInt
Metrics: PA-MPJPE PA-MPVPE F@5mm F@15mm AUC_J AUC_V PCK
Baselines: HaMeR HaMeR-170k MeshGraphormer METRO HandOccNet FrankMocap AMVUR MobRecon HHMR SimpleHand
Builds on: VMamba HaMeR Mamba MANO Graformer
Closest papers by meaning
- Reconstructing Hands in 3D with Transformers (CVPR 2024 · similarity 0.88)
- SiMA-Hand: Boosting 3D Hand-Mesh Reconstruction by Single-to-Multi-View Adaptation (AAAI 2024 · similarity 0.87)
- HOLD: Category-agnostic 3D Reconstruction of Interacting Hands and Objects from Video (CVPR 2024 · similarity 0.86)
- SRHand: Super-Resolving Hand Images and 3D Shapes via View/Pose-aware Neural Image Representations and Explicit Meshes (NeurIPS 2025 · similarity 0.86)
- Fine-Grained Multi-View Hand Reconstruction Using Inverse Rendering (AAAI 2024 · similarity 0.86)
- HandBooster: Boosting 3D Hand-Mesh Reconstruction by Conditional Synthesis and Sampling of Hand-Object Interactions (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.