Research map / 3D vision / Depth and stereo
MVSMamba: Multi-View Stereo with State Space Model
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 MVSMamba, the first Mamba-based multi-view stereo network, using a reference-centered dynamic scanning Mamba module to aggregate global, omnidirectional multi-view features efficiently.
Problem
Transformer-based multi-view stereo (MVS) methods have quadratic attention complexity, making it hard to balance feature-matching performance and efficiency.
Approach
Concatenates source features around the reference feature (top/bottom/left/right) and applies skip scanning with dynamically shifted starting coordinates per source view, so Mamba blocks capture reference-to-source and omnidirectional multi-view dependencies at linear cost.
Key results
- On DTU, MVSMamba* (high-res) reaches overall error 0.280 mm, the best among compared methods, versus 0.287 for MVSFormer++ (high-res) and 0.291 for ET-MVSNet.
- On Tanks-and-Temples, MVSMamba achieves a mean F-score of 67.67 on the intermediate set and 43.32 on the advanced set, the best reported among compared methods.
- MVSMamba (low-res) runs at 0.11 s and 2.82 GB GPU memory with 1.31M parameters on 5-view 832x1152 inputs, versus 0.16 s for ET-MVSNet and 0.24 s for MVSFormer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use high-resolution training | MVSMamba on DTU · DTU evaluation set | Overall: 0.287 -> 0.280 | Helped |
| Use MVSMamba instead of MVSFormer++* | DTU evaluation set · DTU evaluation set | Overall: 0.281 -> 0.287 | Helped |
| Use MVSMamba instead of MVSFormer++ | MVS on Tanks-and-Temples · Tanks-and-Temples intermediate | mean F-score: 67.18 -> 67.67 | Helped |
| Use MVSMamba instead of GoMVS | MVS on Tanks-and-Temples · Tanks-and-Temples advanced | mean F-score: 43.07 -> 43.32 | Helped |
| Use MVSMamba instead of MVSFormer++ | MVS on ETH3D · ETH3D training benchmark | F-score: 82.99 -> 81.69 (-1.5%) | Hurt |
| Remove DM module (ablation) | MVSMamba · DTU evaluation set | Overall: 0.287 -> 0.295 | Hurt |
| Remove SDM module (ablation) | MVSMamba · DTU evaluation set | Overall: 0.287 -> 0.289 | Hurt |
| Remove MLP (ablation) | MVSMamba · DTU evaluation set | Overall: 0.287 -> 0.293 | Hurt |
| Replace proposed aggregation with DCN, FMT, or ET (ablation) | MVSMamba · DTU evaluation set | Overall: 0.287 -> 0.295; 0.296; 0.291 | Hurt |
| Replace dynamic scan with VMamba, EVMamba, or JamMa scan (ablation) | MVSMamba · DTU evaluation set | Overall: 0.287 -> 0.291; 0.298; 0.301 | Hurt |
| Increase the number of DM application scales beyond scale 0 (ablation) | MVSMamba · DTU evaluation set | Overall: 0.289 -> 0.295; 0.292; 0.294 | Hurt |
| Extend SDM from scale 1 to scales 1,2 or 1,2,3 (ablation) | MVSMamba with DM at scale 0 · DTU evaluation set | Overall: 0.287 -> 0.293; 0.296 | Hurt |
Surprising findings
- Adding DM or SDM at more FPN scales does not improve performance, which the authors attribute to incomplete exploitation of Mamba across scales.
- Directly applying JamMa's joint scan or EVMamba's skip scan within the same aggregation setting performs worse than VMamba's four-directional scan in some metrics.
Limitations
- Extending DM/SDM modules to more FPN scales yields no further gains, indicating Mamba's potential is not fully exploited.
- Information loss occurs when global features propagate from coarse to fine scales through the FPN.
- On ETH3D, the F-score is 1.5% lower than MVSFormer++ (completeness is notably lower at 76.85 vs 83.88).
Experimental setup
Datasets: DTU BlendedMVS
Benchmarks: DTU Tanks-and-Temples ETH3D
Metrics: Overall distance (mm) accuracy completeness F-score MAE GPU memory (G) runtime (s) parameters (M)
Baselines: CasMVSNet UniMVSNet MVSTER GeoMVSNet DMVSNet GoMVS TransMVSNet ET-MVSNet MVSFormer MVSFormer++ WT-MVSNet CostFormer
Builds on: Mamba VMamba EVMamba MVSNet MVSFormer++
Closest papers by meaning
- MVSFormer++: Revealing the Devil in Transformer's Details for Multi-View Stereo (ICLR 2024 · similarity 0.89)
- MVSAnywhere: Zero-Shot Multi-View Stereo (CVPR 2025 · similarity 0.87)
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.87)
- MobileMamba: Lightweight Multi-Receptive Visual Mamba Network (CVPR 2025 · similarity 0.86)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.86)
- A Global Depth-Range-Free Multi-View Stereo Transformer Network with Pose Embedding (NeurIPS 2024 · similarity 0.85)
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.