Research map / Point clouds / Point cloud classification
Point-to-Spike Residual Learning for Energy-Efficient 3D Point Cloud Classification
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 P2SResLNet, a spiking neural network that classifies 3D point clouds using binary spikes via spatial-aware kernel point spiking neurons and 3D spiking residual blocks, trained directly with surrogate gradients.
Problem
Energy-efficient 3D point cloud classification: standard point cloud deep networks rely on costly MAC operations, limiting use on embedded robotic systems and neuromorphic hardware.
Approach
Combines integrate-fire spiking neurons with kernel point convolution into spatial-aware kernel point spiking neurons, replacing MAC with sparse AC operations, and builds spiking residual blocks (ADD connection) stacked into a point-to-spike network.
Key results
- P2SResLNet reaches 90.6% OA / 89.2% mAcc on ModelNet40 and 81.2% OA / 79.4% mAcc on ScanObjectNN, versus KPConv ANN at 92.1% and 85.3%.
- Theoretical energy analysis shows the spiking network is over 30 times more energy efficient than KPConv for 6,597 input points.
- The ANN-to-SNN converted KPConv-SNN reaches only 70.5% OA on ModelNet40, far below the directly trained P2SResLNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use residual block design in Figure 2(b) instead of Figure 2(a) (ablation) | P2SResLNet · ModelNet40 | OA: 89.4% vs. 90.6% | Helped |
| Use ADD instead of AND connection (ablation) | P2SResLNet · ModelNet40 | OA: 89.2 -> 90.6 | Helped |
| Use ADD instead of IAND connection (ablation) | P2SResLNet · ModelNet40 | OA: 89.0 -> 90.6 | Helped |
| Use Avgpooling-FC head instead of FC-Avgpooling (ablation) | P2SResLNet · ModelNet40 | OA: 90.0 -> 90.6 | Helped |
| Use Avgpooling-FC head instead of IF-FC-Avgpooling (ablation) | P2SResLNet · ModelNet40 | OA: 89.6 -> 90.6 | Helped |
| Use Avgpooling-FC head instead of IF-Avgpooling-FC (ablation) | P2SResLNet · ModelNet40 | OA: 89.6 -> 90.6 | Helped |
| Decode using accumulated membrane potential instead of spiking sequence (ablation) | P2SResLNet · ModelNet40 | OA: 90.4 -> 90.6 | Helped |
| Set time latency T=1 instead of T=4, T=6, or T=8 (ablation) | P2SResLNet · ModelNet40 | OA: 88.7, 88.6, 88.9 -> 90.6 | Helped |
Surprising findings
- Longer spiking time latency (T=4 to 8) did not improve accuracy on 3D point clouds, in contrast to SNN image classification; T=1 performed best.
Limitations
- Accuracy gap to ANN baselines (about 1.5 points on ModelNet40 and 4 points on ScanObjectNN relative to KPConv)
- Energy results are theoretical; first layer and kernel computation still use MAC and were not deployed on neuromorphic hardware
- Time latency larger than 10 is hard to train due to system limits
Experimental setup
Datasets: ModelNet40 ScanObjectNN
Benchmarks: ModelNet40 ScanObjectNN
Metrics: overall accuracy (OA) mean category accuracy (mAcc) theoretical energy consumption
Baselines: PointNet PointNet++ PointCNN KPConv PointStack PointMLP KPConv-SNN
Builds on: KPConv Integrate-Fire neuron ResNet SpikingJelly surrogate gradient
Closest papers by meaning
- Spiking Point Transformer for Point Cloud Classification (AAAI 2025 · similarity 0.92)
- Noise-Injected Spiking Graph Convolution for Energy-Efficient 3D Point Cloud Denoising (AAAI 2025 · similarity 0.91)
- Efficient 3D Recognition with Event-driven Spike Sparse Convolution (AAAI 2025 · similarity 0.91)
- Rethinking Spiking Neural Networks from an Ensemble Learning Perspective (ICLR 2025 · similarity 0.87)
- SpikePoint: An Efficient Point-based Spiking Neural Network for Event Cameras Action Recognition (ICLR 2024 · similarity 0.87)
- Toward Relative Positional Encoding in Spiking Transformers (NeurIPS 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.