Research map / Point clouds / Point cloud classification
Efficient 3D Recognition with Event-driven Spike Sparse Convolution
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 E-3DSNN, a directly trained spiking 3D backbone combining Spike Voxel Coding and Spike Sparse Convolution with membrane-potential residuals, evaluated on ModelNet40, KITTI detection, and SemanticKITTI segmentation.
Problem
SNNs applied to 3D point clouds underperform ANNs and are limited to simple classification; point-based preprocessing obscures geometry and vanilla spike convolution is computationally inefficient on sparse 3D data.
Approach
Encode voxelized point clouds into sparse spike trains via spike voxel coding, and perform spike convolution only at active (spiking) center locations (SSC), with membrane-potential residual connections (MS-ResNet style) in a voxel-based backbone.
Key results
- E-3DSNN reaches 91.7% top-1 accuracy on ModelNet40 with 1.87M parameters, surpassing the best SNN baselines by 3.0% per the abstract.
- On SemanticKITTI validation, E-3DSNN achieves 69.2% mIoU at 6.1 mJ, versus 67.8% for SphereFormer (49.2 mJ) and 67.2% for SpikePointformer.
- On KITTI car detection (val), E-3DSNN obtains 89.6/84.0/78.7 AP (easy/mod/hard) at 3.4 mJ, above the prior SNN SpikePointRCNN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the spiking E-3DSNN with its same-architecture ANN version (ablation) | E-3DSNN · SemanticKITTI val | mIoU: 69.2 -> 69.4 | Helped |
| Add SSC to the SVC-based E-3DSNN (ablation) | E-3DSNN with SVC · SemanticKITTI val | Power (mJ): 29.1 -> 6.1 | Helped |
| Add SSC to the SVC-based E-3DSNN (ablation) | E-3DSNN with SVC · SemanticKITTI val | mIoU: 69.3 -> 69.2 | Hurt |
| Set T×D to 1×2 instead of 1×4 (ablation) | E-3DSNN · SemanticKITTI val | mIoU: 69.2 -> 67.9 | Hurt |
| Set T×D to 1×2 instead of 1×4 (ablation) | E-3DSNN · SemanticKITTI val | Power (mJ): 6.1 -> 6.3 | Helped |
| Set T×D to 2×1 instead of 2×2 (ablation) | E-3DSNN · SemanticKITTI val | mIoU: 68.9 -> 67.1 | Hurt |
| Set T×D to 2×1 instead of 2×2 (ablation) | E-3DSNN · SemanticKITTI val | Power (mJ): 8.1 -> 8.4 | Hurt |
| Increase T from 2 to 4 while D=1 (ablation) | E-3DSNN · SemanticKITTI val | mIoU: +1.4% | Helped |
| Increase T from 2 to 4 while D=1 (ablation) | E-3DSNN · SemanticKITTI val | Power (mJ) | Hurt |
| Increase D from 1 to 2 while T=2 (ablation) | E-3DSNN · SemanticKITTI val | mIoU: 67.1% vs. 68.9% | Helped |
| Increase D from 1 to 2 while T=2 (ablation) | E-3DSNN · SemanticKITTI val | Power (mJ): 8.4mJ vs. 8.1mJ | Helped |
| Increase the number of channels while keeping the number of blocks constant (ablation) | E-3DSNN encoder · SemanticKITTI | mIoU: 1.8% | Helped |
Surprising findings
- The SNN version of E-3DSNN on SemanticKITTI lost only 0.2% mIoU versus its ANN counterpart while being about 8.8x more energy efficient.
Limitations
- Performance on the large-scale NuScenes dataset is slightly below VoxelNext (NDS 64.1 vs 66.6).
- Energy numbers are theoretical estimates based on FLOPs and firing rates rather than measured hardware results.
Experimental setup
Datasets: ModelNet40 KITTI SemanticKITTI NuScenes
Benchmarks: ModelNet40 KITTI SemanticKITTI NuScenes
Metrics: accuracy 3D AP (R11) mIoU NDS power (mJ) parameters (M)
Baselines: PointNet KPConv Pointformer 3DShapeNets 3DVGG-B SpikePointNet SpikingPointNet P2SResLNet PointRCNN PVRCNN SphereFormer SpikePointformer
Builds on: MS-ResNet Integer LIF neuron OpenPCDet VoxelRCNN Pointcept
Closest papers by meaning
- Point-to-Spike Residual Learning for Energy-Efficient 3D Point Cloud Classification (AAAI 2024 · similarity 0.91)
- Noise-Injected Spiking Graph Convolution for Energy-Efficient 3D Point Cloud Denoising (AAAI 2025 · similarity 0.89)
- Spiking Point Transformer for Point Cloud Classification (AAAI 2025 · similarity 0.89)
- Improving the Sparse Structure Learning of Spiking Neural Networks from the View of Compression Efficiency (ICLR 2025 · similarity 0.88)
- SpikePoint: An Efficient Point-based Spiking Neural Network for Event Cameras Action Recognition (ICLR 2024 · similarity 0.87)
- STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks (CVPR 2025 · similarity 0.87)
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.