Research map / Point clouds / Point cloud classification
Noise-Injected Spiking Graph Convolution for Energy-Efficient 3D Point Cloud Denoising
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 noise-injected spiking graph convolutional networks (NI-SGCN), with purely spiking (NI-PSGCN) and hybrid (NI-HSGCN) variants, for energy-efficient 3D point cloud denoising evaluated on PU-Net and PC-Net.
Problem
Energy-efficient 3D point cloud denoising: SNNs are mostly applied to classification, and their regression potential, plus robustness to stochastic noise, is unexplored for point clouds.
Approach
Builds noise-injected integrate-and-fire neurons (Gaussian noise added to membrane input) and integrates them into a dynamic-graph (EdgeConv-style) spiking graph convolution with dense connections, then uses it as the feature extractor of a ScoreDenoise-like score-estimation network, in either fully spiking or hybrid SNN-ANN form.
Key results
- On PU-Net 50K with 2% noise, NI-HSGCN gets CD 1.798 and P2M 1.147 (x1e-4), versus ScoreDenoise 1.295/0.842 and Pointfilter 1.175/0.709 among the ANN baselines.
- Theoretical energy per forward pass: ANN ScoreDenoise 3.01e9 pJ, NI-PSGCN 2.36e8 pJ (12.75x reduction), NI-HSGCN 7.65e8 pJ (3.94x).
- Mean pooling beats max pooling in the spiking setting (CD 1.798 vs 2.199 on PU-Net).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the hybrid architecture with ANN-based score estimation rather than the purely spiking architecture (ablation) | NI-SGCN · PU-Net, 50K points, 2% noise | CD: 1.926 -> 1.798; reduction of 0.128 | Helped |
| use the hybrid architecture with ANN-based score estimation rather than the purely spiking architecture (ablation) | NI-SGCN · PU-Net, 50K points, 2% noise | P2M: 1.211 -> 1.147; reduction of 0.064 | Helped |
| use mean pooling instead of max pooling (ablation) | NI-SGCN · PU-Net, 50K points, 2% noise | CD: 2.199 -> 1.798; reduces CD by 0.401 | Helped |
| use mean pooling instead of max pooling (ablation) | NI-SGCN · PU-Net, 50K points, 2% noise | P2M: 1.495 -> 1.147; reduces P2M by 0.348 | Helped |
| use noise-injected IF neurons rather than deterministic IF neurons (ablation) | NI-SGCN · PU-Net, 50K points, 2% noise | CD: 2.057 -> 1.798 | Helped |
| use noise-injected IF neurons rather than LIF neurons (ablation) | NI-SGCN · PU-Net, 50K points, 2% noise | CD: 1.889 -> 1.798 | Helped |
| use noise-injected IF neurons rather than noise-injected LIF neurons (ablation) | NI-SGCN · PU-Net, 50K points, 2% noise | CD: 1.848 -> 1.798 | Helped |
| use noise-injected IF neurons rather than QIF neurons (ablation) | NI-SGCN · PU-Net, 50K points, 2% noise | CD: 2.176 -> 1.798 | Helped |
| use noise-injected IF neurons rather than KLIF neurons (ablation) | NI-SGCN · PU-Net, 50K points, 2% noise | CD: 2.469 -> 1.798 | Helped |
| increase latency from 1 to 4 time steps (ablation) | NI-SGCN · PU-Net, 50K points, 2% noise | CD: 5.524 -> 1.798 | Helped |
| increase latency from 4 to 8 time steps (ablation) | NI-SGCN · PU-Net, 50K points, 2% noise | CD: 1.798 -> 1.815 | Hurt |
| increase latency from 1 to 4 time steps (ablation) | NI-SGCN · PU-Net, 50K points, 2% noise | P2M: 4.419 -> 1.147 | Helped |
Surprising findings
- Under the reported tables, the ANN baselines ScoreDenoise and Pointfilter beat the SNN variants on most PU-Net settings, so the SNNs trade some accuracy for energy.
Limitations
- Energy results are theoretical operation-count estimates, not measured on neuromorphic hardware.
- SNN variants show higher CD/P2M than the best ANN baselines on most settings.
- Evaluation uses only PU-Net and PC-Net with Gaussian noise at 1-2%.
Experimental setup
Datasets: PU-Net PC-Net
Benchmarks: PU-Net PC-Net
Metrics: Chamfer Distance Point-to-Mesh Distance theoretical energy (pJ)
Baselines: PCN ScoreDenoise DMRDenoise Pointfilter QIF KLIF
Builds on: ScoreDenoise EdgeConv SpikingJelly
Closest papers by meaning
- Point-to-Spike Residual Learning for Energy-Efficient 3D Point Cloud Classification (AAAI 2024 · similarity 0.91)
- Spiking Point Transformer for Point Cloud Classification (AAAI 2025 · similarity 0.91)
- Efficient 3D Recognition with Event-driven Spike Sparse Convolution (AAAI 2025 · similarity 0.89)
- Rethinking Spiking Neural Networks from an Ensemble Learning Perspective (ICLR 2025 · similarity 0.87)
- S$^2$NN: Sub-bit Spiking Neural Networks (NeurIPS 2025 · similarity 0.86)
- SGNNBench: A Holistic Evaluation of Spiking Graph Neural Networks on Large-scale Graphs (NeurIPS 2026 · 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.