Research map / Point clouds / Point cloud classification
PointMamba: A Simple State Space Model for Point Cloud Analysis
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 PointMamba, a plain non-hierarchical Mamba encoder for point clouds that tokenizes points along Hilbert and Trans-Hilbert space-filling curves with an order indicator, pre-trained via serialization-based masked modeling, achieving linear complexity and competitive accuracy.
Problem
Transformers for point cloud analysis have quadratic attention complexity, and it is unclear whether state space models like Mamba, which are unidirectional and designed for 1D sequences, can model unstructured 3D point clouds effectively.
Approach
Serialize unordered point clouds with Hilbert and Trans-Hilbert space-filling curves into two token groups, distinguish them with a learned order indicator, and feed them into a vanilla Mamba encoder so the second sequence receives global context from the first; pre-train with random-curve serialization masked reconstruction.
Key results
- On ScanObjectNN PB-T50-RS, PointMamba reaches 89.31% vs Point-MAE 89.04% (with rotation augmentation) using 12.3M params and 3.1G FLOPs versus 22.1M and 4.8G.
- Beyond 32,768 tokens, PointMamba is reported as 30.2x faster in inference, 24.9x lower GPU memory, and 5.2x fewer FLOPs than Point-MAE.
- On ModelNet40 PointMamba surpasses Point-MAE by 0.4% and PointGPT-S by 0.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Hilbert and Trans-Hilbert scanning instead of random serialization (ablation) | PointMamba · ScanObjectNN OBJ-BG | overall accuracy: +1.20% | Helped |
| Use Hilbert and Trans-Hilbert scanning instead of random serialization (ablation) | PointMamba · ScanObjectNN OBJ-ONLY | overall accuracy: +1.73% | Helped |
| Add distinct order indicators for Hilbert and Trans-Hilbert sequences (ablation) | PointMamba · ScanObjectNN OBJ-BG | overall accuracy: +1.20% | Helped |
| Add distinct order indicators for Hilbert and Trans-Hilbert sequences (ablation) | PointMamba · ScanObjectNN OBJ-ONLY | overall accuracy: +1.03% | Helped |
| Use the same order indicator for both scanning sequences (ablation) | PointMamba · ScanObjectNN | overall accuracy | No clear effect |
| Add Selective SSM instead of using the identity setting (ablation) | PointMamba · ScanObjectNN | overall accuracy | Helped |
| Replace Selective SSM with masked self-attention (ablation) | PointMamba · ScanObjectNN | overall accuracy | Hurt |
| Replace Selective SSM with MLP (ablation) | PointMamba · ScanObjectNN | overall accuracy | Mixed |
Surprising findings
- Replacing the selective SSM with masked self-attention performed worse than the identity (no SSM) setting, attributed to difficulty combining attention with Mamba's gated MLP.
- Masked self-attention and the default Mamba block behave differently on point clouds, contrary to MambaOut's claim that SSM hurts image classification, here SSM helped.
Limitations
- Focuses only on point cloud analysis; not a unified 3D foundation model for detection or other 3D tasks.
- Uses only point clouds as training data without 2D image or language knowledge.
- Masked self-attention variant underperforms, leaving the attention-SSM combination unexplored.
Experimental setup
Datasets: ShapeNetCore ScanObjectNN ModelNet40 ShapeNetPart
Benchmarks: ScanObjectNN ModelNet40 ShapeNetPart
Metrics: overall accuracy mIoU few-shot accuracy parameters FLOPs GPU memory inference speed
Baselines: Point-MAE PointGPT-S Point-M2AE Point-BERT MaskPoint ACT ReCon I2P-MAE PointNeXt OctFormer Point Cloud Mamba Mamba3D
Builds on: Mamba Structured State Space Sequence Model (S4) Point-MAE Hilbert space-filling curve
Closest papers by meaning
- Point Cloud Mamba: Point Cloud Learning via State Space Model (AAAI 2025 · similarity 0.93)
- Pamba: Enhancing Global Interaction in Point Clouds via State Space Model (AAAI 2025 · similarity 0.91)
- Spectral Informed Mamba for Robust Point Cloud Processing (CVPR 2025 · similarity 0.91)
- LCM: Locally Constrained Compact Point Cloud Model for Masked Point Modeling (NeurIPS 2024 · similarity 0.90)
- ZigzagPointMamba: Spatial-Semantic Mamba for Point Cloud Understanding (NeurIPS 2025 · similarity 0.89)
- PointRWKV: Efficient RWKV-Like Model for Hierarchical Point Cloud Learning (AAAI 2025 · similarity 0.89)
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.