Research map / Point clouds / Point cloud classification
Pamba: Enhancing Global Interaction in Point Clouds 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.
Pamba applies Mamba (SSM) to whole-scene 3D point cloud semantic segmentation on top of PTv3, adding multi-path (hz) serialization and a ConvMamba block with bidirectional SSM and sparse convolutions.
Problem
Transformers for 3D point cloud segmentation have quadratic complexity, limiting how many points can be processed at once and hindering long-range interactions between objects in a scene.
Approach
Process the entire point cloud as a single sequence with bidirectional Mamba (shared-parameter forward/backward SSM), preceded by sparse submanifold convolutions for local geometry, and randomly assign Hilbert/z-order hybrid (hz, hz-swap) serialization curves across blocks to handle permutation sensitivity.
Key results
- Pamba reaches 77.6 mIoU on ScanNet v2 val, 36.3 on ScanNet200 val (+1.1 over prior SOTA), 73.5 on S3DIS Area 5 and 80.4 on nuScenes val, matching or slightly exceeding PTv3.
- Ablation: removing Mamba drops ScanNet200 val mIoU from 36.3 to 26.1; unidirectional Mamba gives 33.2.
- Pamba is slower than PTv3 at inference (179ms vs 108ms) with slightly higher inference memory (4.8G vs 4.5G), while training memory matches PTv3 at 5.2G.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Pamba instead of PTv3 | Point cloud semantic segmentation · ScanNet200 Val | mIoU: 35.2 -> 36.3 | Helped |
| Use Pamba instead of PTv3 | Point cloud semantic segmentation · ScanNet v2 Val | mIoU: 77.5 -> 77.6 | Helped |
| Use Pamba instead of PTv3 | Point cloud semantic segmentation · nuScenes Val | mIoU: 80.4 -> 80.4 | No clear effect |
| Use Pamba instead of PTv3 | Point cloud semantic segmentation · S3DIS Area 5 | mIoU: 73.4 -> 73.5 | Helped |
| Use Pamba instead of PTv3 | ScanNet200 · ScanNet200 | inference latency: 108ms -> 179ms | Hurt |
| Remove all Mamba modules (ablation) | Pamba on ScanNet200 · ScanNet200 Val | mIoU: 36.3 -> 26.1 | Hurt |
| Replace bidirectional Mamba with unidirectional Mamba (ablation) | Pamba on ScanNet200 · ScanNet200 Val | mIoU: 36.3 -> 33.2 | Hurt |
| Use different parameters for forward and backward SSM (ablation) | Pamba on ScanNet200 · ScanNet200 Val | mIoU: 36.3 -> 35.7 | Hurt |
| Add sparse-convolution local aggregation (ablation) | Pamba on ScanNet200 · ScanNet200 Val | mIoU: 31.9 -> 36.3 | Helped |
| Use 1 rather than 2 sparse convolutions for local aggregation (ablation) | Pamba on ScanNet200 · ScanNet200 Val | mIoU: 36.3 -> 35.9 | Hurt |
| Use 3 rather than 2 sparse convolutions for local aggregation (ablation) | Pamba on ScanNet200 · ScanNet200 Val | mIoU: 36.3 -> 36.0 | Hurt |
| Use 4 rather than 2 sparse convolutions for local aggregation (ablation) | Pamba on ScanNet200 · ScanNet200 Val | mIoU: 36.3 -> 35.4 | Hurt |
Surprising findings
- Inference latency of Pamba (179ms) is notably higher than PTv3 (108ms), attributed to serial per-point computation in the SSM scan.
Limitations
- Slower inference latency than transformer-based PTv3 and OctFormer due to serial SSM computation.
- Evaluated only on semantic segmentation; instance segmentation and detection not tested.
- Matches rather than clearly beats PTv3 on several datasets.
Experimental setup
Datasets: ScanNet v2 ScanNet200 S3DIS nuScenes
Benchmarks: ScanNet v2 ScanNet200 S3DIS Area 5 nuScenes
Metrics: mIoU training latency training memory inference latency inference memory
Baselines: PTv3 PTv2 OctFormer Swin3D PCM SphereFormer Swim3D PointNeXt KPConv StratifiedFormer
Builds on: Point Transformer V3 Mamba OctFormer Hilbert curve z-order curve
Closest papers by meaning
- PointMamba: A Simple State Space Model for Point Cloud Analysis (NeurIPS 2024 · similarity 0.91)
- Point Cloud Mamba: Point Cloud Learning via State Space Model (AAAI 2025 · similarity 0.91)
- Spectral Informed Mamba for Robust Point Cloud Processing (CVPR 2025 · similarity 0.89)
- ZigzagPointMamba: Spatial-Semantic Mamba for Point Cloud Understanding (NeurIPS 2025 · similarity 0.88)
- PMA: Towards Parameter-Efficient Point Cloud Understanding via Point Mamba Adapter (CVPR 2025 · similarity 0.86)
- Point-SAM: Promptable 3D Segmentation Model for Point Clouds (ICLR 2025 · 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.