Research map / Point clouds / Point cloud classification
Point Cloud Mamba: Point Cloud Learning via 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 Point Cloud Mamba (PCM), a local-plus-global point cloud network that serializes point clouds with six Consistent Traverse Serialization variants, adds order prompts and coordinate-based positional embeddings, and reports SOTA on ScanObjectNN, ModelNet40, ShapeNetPart and S3DIS.
Problem
Point cloud analysis with state space models: Mamba is designed for causal 1-D sequences, so converting unordered, irregular 3-D point sets into sequences that preserve spatial neighborhoods while achieving global modeling at linear cost remains open; prior Mamba point cloud work (PointMamba) lags p
Approach
Serialize voxelized point clouds into 1-D sequences via a grid-code sort in which spatially adjacent points stay adjacent, use six axis-permuted variants across Mamba layers, add learnable order prompts marking each serialization, and embed points with a linear map of their xyz coordinates, all within a PointMLP-style local-plus-global hierarchy.
Key results
- PCM reaches 88.1% OA and 86.6% mAcc on ScanObjectNN (PB_T50_RS), beating PointNeXt (87.7%) and PTv3 (86.4%).
- With additional DeLA local blocks, PCM-Tiny reaches 79.6 mIoU on S3DIS Area 5, versus 74.1 mIoU for DeLA-free PCM-Tiny and 74.1 mIoU for DeLA.
- PCM reaches 87.0 instance mIoU on ShapeNetPart, surpassing PointNeXt (86.5 ins. mIoU reported in the table).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace naive Mamba architecture with PCM design | Mamba point cloud network · ScanObjectNN | OA: 2.8 OA | Helped |
| Use multiple serialization strategies instead of a single strategy (ablation) | PCM · ScanObjectNN | OA: 0.42 OA | Helped |
| Use xyz, yzx, and zxy CTS variants together (ablation) | PCM · ScanObjectNN | OA: 0.17 OA | Helped |
| Combine all six CTS serialization variants (ablation) | PCM · ScanObjectNN | OA: 0.39 OA | Helped |
| Reduce channel size from 768 to 384 (ablation) | PCM · ScanObjectNN | OA: 1.12 OA | Hurt |
| Reduce channel size to 96 (ablation) | PCM · ScanObjectNN | OA: 2.56 OA | Hurt |
| Replace RoPE with spatial coordinate Linear positional mapping (ablation) | PCM · ScanObjectNN | OA: 0.37 OA | Helped |
| Replace Linear positional mapping with an MLP (ablation) | PCM · ScanObjectNN | OA | No clear effect |
| Use a separate Linear positional mapping for each Mamba layer (ablation) | PCM · ScanObjectNN | OA | Hurt |
| Add six order prompts instead of using none (ablation) | PCM · ScanObjectNN | OA: 0.83 OA | Helped |
| Increase order prompts beyond six (ablation) | PCM · ScanObjectNN | OA: 87.47; 87.13 | Hurt |
| Add local neighborhood feature extraction using four neighboring points (ablation) | PCM · ScanObjectNN | OA: 5.34 OA | Helped |
Surprising findings
- Random flattening of point clouds into sequences significantly underperforms ordered serialization (84.1 vs 86.7 OA in the text).
- Adding PTv3's local DeLA extractor did not improve PTv3, which the authors attribute to windowed attention limits.
Limitations
- Scan-based Mamba training scales poorly to large point clouds (>=100k points), forcing cropping on S3DIS and a train/test gap.
- Performance gains on ModelNet40 are small and within run-to-run variance, with the 94.1 PointMLP figure requiring best-of-many selection.
- Throughput is lower than PointMLP despite fewer parameters and FLOPs.
Experimental setup
Datasets: ScanObjectNN ModelNet40 ShapeNetPart S3DIS ScanNet
Benchmarks: ScanObjectNN ModelNet40 ShapeNetPart S3DIS ScanNet
Metrics: OA mAcc instance mIoU class mIoU mIoU params FLOPs throughput
Baselines: PointNeXt PointMLP PTv3 PointMamba DeLA Point Transformer PTv2 Point-MAE
Builds on: Mamba PointMLP PointNeXt DeLA S4
Closest papers by meaning
- PointMamba: A Simple State Space Model for Point Cloud Analysis (NeurIPS 2024 · similarity 0.93)
- Spectral Informed Mamba for Robust Point Cloud Processing (CVPR 2025 · similarity 0.91)
- PMA: Towards Parameter-Efficient Point Cloud Understanding via Point Mamba Adapter (CVPR 2025 · similarity 0.91)
- Pamba: Enhancing Global Interaction in Point Clouds via State Space Model (AAAI 2025 · similarity 0.91)
- ZigzagPointMamba: Spatial-Semantic Mamba for Point Cloud Understanding (NeurIPS 2025 · similarity 0.89)
- LCM: Locally Constrained Compact Point Cloud Model for Masked Point Modeling (NeurIPS 2024 · similarity 0.88)
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.