Research map / Point clouds / Point cloud classification
FlowMamba: Learning Point Cloud Scene Flow with Global Motion Propagation
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 FlowMamba, a point cloud scene flow network whose iterative update module uses bidirectional Mamba blocks with a feature-induced point ordering to propagate global motion, reaching millimeter-level EPE3D on FlyingThings3D and KITTI.
Problem
Scene flow estimation from point clouds struggles in ill-posed regions (flat areas, slender structures, occlusions) where local evidence is insufficient, and Transformers are costly for global context.
Approach
An SSM-based iterative update unit (ISU) built from bidirectional Mamba blocks propagates global motion and gates the fusion with prior hidden states, while a feature-induced ordering (FIO) sorts points by a learned score from context, correlation and hidden features so that the sequence respects spatial continuity.
Key results
- On FlyingThings3D (no occlusion), EPE3D 0.0089 vs. DifFlow3D 0.0114, a 21.9% reduction; on KITTI (no occlusion) EPE3D 0.0062, a 20.5% reduction.
- With occlusion, EPE3D 0.034 on FT3D and 0.028 on KITTI, reductions of 20.9% and 9.6% over DifFlow3D.
- Plugging ISU into PV-RAFT, PT-FlowNet and MSBRN reduces EPE3D on FlyingThings3D by 28.2%, 20.4% and 28.6% respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FlowMamba instead of DifFlow3D | FlowMamba · FlyingThings3D without occlusion | EPE3D: 0.0114 -> 0.0089 | Helped |
| Use FlowMamba instead of DifFlow3D | FlowMamba · KITTI without occlusion | EPE3D: 0.0078 -> 0.0062 | Helped |
| Use FlowMamba instead of DifFlow3D | FlowMamba · FlyingThings3D with occlusion | EPE3D: 0.043 -> 0.034 | Helped |
| Use FlowMamba instead of DifFlow3D | FlowMamba · KITTI with occlusion | EPE3D: 0.031 -> 0.028 | Helped |
| Replace the baseline GRU with ISU | PV-RAFT · FlyingThings3D without occlusion | EPE3D: 0.0461 -> 0.0331 | Helped |
| Replace the baseline GRU with ISU | PT-FlowNet · KITTI without occlusion | EPE3D: 0.0224 -> 0.0209 | Helped |
| Replace Conv GRU with unidirectional Mamba (ablation) | FlowMamba · FlyingThings3D | EPE3D: 0.0134 -> 0.0141 | Hurt |
| Replace Conv GRU with bidirectional Mamba (ablation) | FlowMamba · FlyingThings3D | EPE3D: 0.0134 -> 0.0127 | Helped |
| Replace bidirectional Mamba with ISU (ablation) | FlowMamba · FlyingThings3D | EPE3D: 0.0127 -> 0.0101 | Helped |
| Add FIO to ISU (ablation) | FlowMamba · FlyingThings3D | EPE3D: 0.0101 -> 0.0089 | Helped |
| Add context feature to FIO (ablation) | FlowMamba · FlyingThings3D | EPE3D: 0.0096 -> 0.0089 | Helped |
| Add correlation feature to FIO (ablation) | FlowMamba · FlyingThings3D | EPE3D: 0.0095 -> 0.0089 | Helped |
Surprising findings
- Bidirectional Mamba alone gives only marginal gains over unidirectional Mamba, which the authors attribute to global noise propagation.
- FlowMamba with 2 iterations matches MSBRN with 4 iterations on FlyingThings3D.
Limitations
- Relies on FlyingThings3D supervised training; generalization to KITTI is evaluated without fine-tuning only.
- Ordering and global propagation add runtime relative to the baseline (304ms vs 283ms).
Experimental setup
Datasets: FlyingThings3D
Benchmarks: FlyingThings3D KITTI
Metrics: EPE3D Acc3DS Acc3DR Outliers EPE2D Acc2D
Baselines: FlowNet3D HPLFlowNet PointPWC-Net HALFlow FLOT PV-RAFT RCP 3DFlow Bi-PointFlowNet PT-FlowNet IHNet MSBRN
Builds on: MSBRN Mamba S4 PointConv PV-RAFT
Closest papers by meaning
- DeltaFlow: An Efficient Multi-frame Scene Flow Estimation Method (NeurIPS 2025 · similarity 0.86)
- ZeroFlow: Scalable Scene Flow via Distillation (ICLR 2024 · similarity 0.86)
- Zero-Shot Monocular Scene Flow Estimation in the Wild (CVPR 2025 · similarity 0.86)
- OAMaskFlow: Occlusion-Aware Motion Mask for Scene Flow (AAAI 2025 · similarity 0.86)
- MESSENGER: Memory-Enhanced Sequential Scene Flow Estimation via Autoregressive Next-Frame Forecasting (NeurIPS 2026 · similarity 0.86)
- ICP-Flow: LiDAR Scene Flow Estimation with ICP (CVPR 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.