Research map / Point clouds / Point cloud classification

Spiking Point Transformer for Point Cloud Classification

AAAI 2025 · Point clouds › Point cloud classification · approach: Efficient point sampling architectures · 13 citations (OpenAlex)

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 Spiking Point Transformer (SPT), the first transformer-based spiking neural network for point cloud classification, using queue-driven sampling direct encoding (Q-SDE) and a hybrid-dynamics integrate-and-fire neuron (HD-IF).

Problem

Transformer-based spiking neural networks have been explored for 2D images but not for 3D point clouds; direct encoding of point clouds over many time steps is computationally expensive and sparse point clouds gain little from longer time steps.

Approach

Q-SDE splits point subsets across time steps via FIFO queue-driven FPS sampling instead of repeating the full point cloud at each step, and HD-IF gates a mixture of IF, LIF, EIF and PLIF neuron dynamics (Top-2 at inference) to selectively activate neurons.

Key results

• SPT reaches 91.43% OA on ModelNet40 and 94.76% OA on ModelNet10 at 4 time steps, the best among SNN methods reported.
• On ScanObjectNN, SPT attains 78.03% OA without voting (3.57% over P2SResLNet-B) and 82.23% with voting.
• Estimated energy is 3.0 mJ at 1 time step (28.2x less than the Point Transformer ANN) and 13.3 mJ at 4 time steps (about 6.4x less).

Measured effects reported in the paper

Surprising findings

• Unlike prior point-cloud SNNs, accuracy generally improves with more time steps in SPT.
• Too many support points (Q-SDE1024/768) can hurt accuracy at longer time steps, so Q-SDE512 performs best at 4 steps.

Limitations

• Energy figures are theoretical estimates based on MAC/AC operation counts rather than measured on neuromorphic hardware.
• SPT is slightly below Point Transformer on ModelNet40 without the SNN constraint, and Q-SDE768 underperforms Q-SDE512 at 4 steps.
• Evaluated only on classification; segmentation and detection are left for future work.

Experimental setup

Datasets: ModelNet10 ModelNet40 ScanObjectNN

Benchmarks: ModelNet10 ModelNet40 ScanObjectNN

Metrics: overall accuracy mean class accuracy energy consumption (mJ) runtime memory

Baselines: PointNet PointNet++ Point Transformer PointMLP KPConv-SNN Spiking Pointnet P2SResLNet-B

Builds on: Point Transformer Spikformer Spiking PointNet SpikingJelly

Closest papers by meaning

• Point-to-Spike Residual Learning for Energy-Efficient 3D Point Cloud Classification (AAAI 2024 · 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 0.89)
• Spiking Transformer: Introducing Accurate Addition-Only Spiking Self-Attention for Transformer (CVPR 2025 · similarity 0.89)
• STEP: A Unified Spiking Transformer Evaluation Platform for Fair and Reproducible Benchmarking (NeurIPS 2025 · similarity 0.88)
• Rethinking Spiking Neural Networks from an Ensemble Learning Perspective (ICLR 2025 · similarity 0.88)