Research map / Point clouds / Point cloud classification
ZigzagPointMamba: Spatial-Semantic Mamba for Point Cloud Understanding
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 ZigzagPointMamba, a PointMamba-based masked autoencoder that uses a 3D zigzag scan ordering of point tokens and a Semantic-Siamese Masking Strategy that masks redundant, semantically similar tokens before random masking.
Problem
PointMamba-based self-supervised point cloud methods use conventional scan orders (random, Hilbert, Z-order) that disrupt spatial continuity, and random masking that fails to capture global semantic dependencies.
Approach
A multi-plane zigzag scan path orders point tokens so spatially adjacent points stay adjacent in the sequence, combined with a masking scheme that first masks tokens with high cosine-similarity redundancy (SMS) and then applies random masking to the rest.
Key results
- Reports 92.10% on OBJ-ONLY and 88.65% on PB-T50-RS for ScanObjectNN, versus PointMamba baseline 90.88% and 87.93%.
- ModelNet40 overall accuracy 93.15% vs PointMamba 92.75% (no voting).
- ShapeNetPart Cls.mIoU 84.16% vs PointMamba 82.57%; Inst.mIoU 85.78% vs 85.28%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ZigzagPointMamba instead of PointMamba | PointMamba · ModelNet40 | overall accuracy: 92.75 -> 93.15 | Helped |
| Use ZigzagPointMamba instead of PointMamba | PointMamba · ShapeNetPart | class mIoU: 82.57 -> 84.16 | Helped |
| Use ZigzagPointMamba instead of PointMamba | PointMamba · ScanObjectNN OBJ-ONLY | accuracy: 90.88 -> 92.10 | Helped |
| Use ZigzagPointMamba instead of PointMamba | PointMamba · ScanObjectNN OBJ-BG | accuracy: 93.96 -> 94.15 | Helped |
| Use ZigzagPointMamba instead of PointMamba | PointMamba · ScanObjectNN PB-T50-RS | accuracy: 87.93 -> 88.65 | Helped |
| Set SMS ratio to 0.8 rather than the other tested ratios (ablation) | ZigzagPointMamba · ScanObjectNN OBJ-ONLY | accuracy: 90.71, 91.57, 91.22, 92.08, 91.91 | Helped |
| Set SMS ratio to 0.8 rather than the other tested ratios (ablation) | ZigzagPointMamba · ScanObjectNN PB-T50-RS | accuracy: 87.99, 87.68, 88.45, 88.65, 88.36 | Helped |
| Replace zigzag scan with random scan (ablation) | ZigzagPointMamba · ScanObjectNN OBJ-ONLY | accuracy: 92.10 -> 92.60 | Mixed |
| Replace zigzag scan with random scan (ablation) | ZigzagPointMamba · ScanObjectNN PB-T58-RS | accuracy: 88.65 -> 90.18 | Helped |
| Replace zigzag scan with Z-order and Trans-Z-order (ablation) | ZigzagPointMamba · ScanObjectNN OBJ-ONLY | accuracy: 92.10 -> 93.29 | Helped |
| Replace zigzag scan with Hilbert and Z-order (ablation) | ZigzagPointMamba · ScanObjectNN PB-T58-RS | accuracy: 88.65 -> 90.88 | Helped |
| Replace zigzag scan with Trans-Hilert and Trans-Z-order (ablation) | ZigzagPointMamba · ScanObjectNN PB-T58-RS | accuracy: 88.65 -> 91.91 | Helped |
Surprising findings
- The zigzag scan path alone does not clearly outperform simple baselines: on PB-T58-RS, Trans-Hilbert with Trans-Z-order (91.91%) beats the zigzag path (88.65% on PB-T50-RS in the table), and the authors concede the scan curve results are not particularly outstanding.
- Reported ScanObjectNN OBJ-BG result (94.15%) ties ReCon, and the paper's own ablation numbers are inconsistent across tables (e.g., 92.08% vs 92.10% on OBJ-ONLY; PB-T50-RS vs PB-T58-RS naming).
Limitations
- Fixed SMS similarity threshold (0.8) may not generalize across point cloud distributions.
- Unidirectional state space modeling limits extension to temporal or video point cloud data.
- Zigzag scan path gives only modest or no gains over alternative scan curves in ablation.
Experimental setup
Datasets: ShapeNet ModelNet40 ScanObjectNN ShapeNetPart ModelNetFewShot
Benchmarks: ScanObjectNN ModelNet40 ShapeNetPart ModelNetFewShot
Metrics: overall accuracy instance mIoU class mIoU few-shot accuracy
Baselines: Point-BERT MaskPoint PointMAE PointM2AE ACT ReCon GeoMask3D PointMamba PointGPT-S
Builds on: PointMamba Point-MAE Mamba
Closest papers by meaning
- PointMamba: A Simple State Space Model for Point Cloud Analysis (NeurIPS 2024 · similarity 0.89)
- Spectral Informed Mamba for Robust Point Cloud Processing (CVPR 2025 · similarity 0.89)
- Point Cloud Mamba: Point Cloud Learning via State Space Model (AAAI 2025 · similarity 0.89)
- Pamba: Enhancing Global Interaction in Point Clouds via State Space Model (AAAI 2025 · similarity 0.88)
- LCM: Locally Constrained Compact Point Cloud Model for Masked Point Modeling (NeurIPS 2024 · similarity 0.86)
- PointDGMamba: Domain Generalization of Point Cloud Classification via Generalized State Space Model (AAAI 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.