Research map / Point clouds / Point cloud classification
PointRWKV: Efficient RWKV-Like Model for Hierarchical Point Cloud Learning
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 PointRWKV, a linear-complexity hierarchical point cloud encoder built from RWKV-style blocks with a parallel fixed-radius graph branch, pre-trained on ShapeNet with masked point modeling.
Problem
Transformers for point cloud learning have quadratic attention complexity, limiting scaling to long point sequences under limited compute; existing Mamba-based point models are unidirectional.
Approach
Adapts RWKV to point clouds with bidirectional quadratic expansion, dynamic bidirectional global attention with data-dependent decay, and a parallel local graph-based merging branch with a graph stabilizer, organized as a multi-scale U-Net-like hierarchy.
Key results
- After pre-training, PointRWKV reaches 93.63% overall accuracy on ScanObjectNN PB-T50-RS and 96.89% OA on ModelNet40, versus Mamba3D's 95.1% on ModelNet40.
- On ShapeNetPart, PointRWKV achieves 90.26% instance mIoU, 3.16 points above prior methods, with 15.2M params and 6.5G FLOPs.
- Reports roughly 42% FLOPs savings and 15.4x GFLOPs reduction versus a transformer baseline at long sequence lengths.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PointRWKV instead of Mamba3D | Point cloud classification · ScanObjectNN PB-T50-RS | accuracy: 88.97 -> 93.63 | Helped |
| Use PointRWKV instead of Mamba3D | Point cloud classification · ModelNet40 | OA: 95.1 -> 96.89 | Helped |
| Use PointRWKV instead of Point-M2AE | Few-shot classification · ModelNet40 5-way 10-shot | accuracy: +1.1 | Helped |
| Use PointRWKV instead of Point-M2AE | Few-shot classification · ModelNet40 5-way 20-shot | accuracy: +0.9 | Helped |
| Use PointRWKV instead of Point-M2AE | Few-shot classification · ModelNet40 10-way 10-shot | accuracy: +2.5 | Helped |
| Use PointRWKV instead of Point-M2AE | Few-shot classification · ModelNet40 10-way 20-shot | accuracy: +1.5 | Helped |
| Use PointRWKV instead of PCM | Part segmentation · ShapeNetPart | instance mIoU: 87.1 -> 90.26 | Helped |
| Add BQE (ablation) | PointRWKV baseline · ScanObjectNN PB-T50-RS | OA: 89.6 -> 90.4 | Helped |
| Add bidirectional attention (ablation) | PointRWKV with BQE · ScanObjectNN PB-T50-RS | OA: 90.4 -> 91.6 | Helped |
| Add multi-scale point cloud learning (ablation) | PointRWKV with bidirectional attention · ScanObjectNN PB-T50-RS | OA: 91.6 -> 92.8 | Helped |
| Add LGM without graph stabilizer (ablation) | PointRWKV with multi-scale learning · ScanObjectNN PB-T50-RS | OA: 92.8 -> 93.0 | Helped |
| Add graph stabilizer to LGM (ablation) | PointRWKV with LGM · ScanObjectNN PB-T50-RS | OA: 93.0 -> 93.6 | Helped |
Limitations
- Still relies on a pre-training scheme; authors say more elegant methods are needed.
- Only an initial attempt at applying RWKV to 3D point clouds.
Experimental setup
Datasets: ShapeNet ModelNet40 ScanObjectNN ShapeNetPart S3DIS
Benchmarks: ScanObjectNN ModelNet40 ShapeNetPart S3DIS
Metrics: overall accuracy mean accuracy instance mIoU class mIoU parameters FLOPs few-shot accuracy
Baselines: Point-MAE Point-M2AE Point-BERT PointMamba Mamba3D PCM PointNeXt PointMLP
Builds on: RWKV Point-MAE PointNet++ Vision-RWKV
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 0.87)
- Point Cloud Mamba: Point Cloud Learning via State Space Model (AAAI 2025 · similarity 0.87)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.86)
- PointDGMamba: Domain Generalization of Point Cloud Classification via Generalized State Space Model (AAAI 2025 · similarity 0.86)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · 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.