Abstract
Spiking neural networks (SNNs), inspired by the spiking computation paradigm of the biological neural systems, have exhibited superior energy efficiency in 2D classification tasks over traditional artificial neural networks (ANNs). However, the regression potential of SNNs has not been well explored, especially in 3D point cloud processing. In this paper, we propose noise-injected spiking graph convolutional networks to leverage the full regression potential of SNNs in 3D point cloud denoising. Specifically, we first emulate the noise-injected neuronal dynamics to build noise-injected spiking neurons. On this basis, we design noise-injected spiking graph convolution for promoting disturbance-aware spiking representation learning on 3D points. Starting from the spiking graph convolution, we build two SNN-based denoising networks. One is a purely spiking graph convolutional network, which achieves low accuracy loss compared with some ANN-based alternatives, while resulting in significantly reduced energy consumption on two benchmark datasets, PU-Net and PC-Net. The other is a hybrid architecture that combines ANN-based learning with a high performance-efficiency trade-off in just a few time steps. Our work lights up SNN's potential for 3D point cloud denoising, injecting new perspectives of exploring the deployment on neuromorphic chips while paving the way for developing energy-efficient 3D data acquisition devices.
Introduction
3D point clouds generated by LiDAR and depth scanners are fundamental to robotic perception but are frequently corrupted by sensor noise. While Artificial Neural Networks (ANNs) provide state-of-the-art denoising, their reliance on power-intensive Multiply-Accumulate (MAC) operations poses a significant challenge for battery-constrained edge devices like drones and AR glasses. Spiking Neural Networks (SNNs) offer a promising alternative by utilizing energy-efficient Accumulate (AC) operations and sparse communication. However, the deterministic nature of standard SNNs often struggles with the high-precision regression required for 3D coordinate prediction. This work introduces NI-SGCN, a noise-injected spiking graph convolutional network designed to bridge this gap.
Methodology
The core innovation lies in the Noise-Injected Integrate-and-Fire (NIIF) neuron. By emulating the stochasticity of biological brains, Gaussian noise is injected into the membrane potential during training, transforming the firing mechanism from deterministic to probabilistic. This approach forces the network to learn robust representations that are less sensitive to input disturbances.
Building upon these neurons, the authors developed Noise-Injected Spiking Graph Convolution (NI-SGC). This operator adapts the "EdgeConv" paradigm to the spiking domain, capturing local geometric structures through sparse spikes. Two architectures were evaluated: NI-PSGCN, a pure SNN for maximum efficiency, and NI-HSGCN, a hybrid model that utilizes SNNs for feature extraction and a lightweight ANN for score estimation.
Experimental Results
The models were benchmarked on the PU-Net and PC-Net datasets across various noise levels. The pure SNN (NI-PSGCN) demonstrated remarkable efficiency, consuming only 2.36 × 10⁸ pJ per forward pass—a 12.75x reduction compared to the ANN-based ScoreDenoise.
In terms of accuracy, the hybrid NI-HSGCN achieved a Chamfer Distance (CD) of 1.798 on the PU-Net dataset (50K points, 2% noise), closely rivaling the ANN baseline (1.295) and outperforming traditional methods like PCN. Qualitative analysis confirms that the proposed method effectively preserves sharp edges and fine details while smoothing noisy surfaces.
Conclusion
This research demonstrates that SNNs can be successfully adapted for high-precision 3D regression tasks. By injecting noise into neuronal dynamics, the proposed NI-SGCN achieves a superior balance between energy efficiency and denoising performance. Future work will focus on closing the remaining accuracy gap and deploying these models on physical neuromorphic hardware to realize real-world power savings.
Summary updated