Introduction
3D point clouds, generated by LiDAR and 3D scanners, are fundamental to robotic perception and autonomous systems. However, these raw data points are often corrupted by noise from sensor imperfections and environmental factors. While traditional Artificial Neural Networks (ANNs) have achieved high accuracy in point cloud denoising, they are computationally intensive, relying on power-hungry multiply-accumulate (MAC) operations. This poses a significant challenge for deployment on battery-constrained edge devices like drones or portable scanners.
Spiking Neural Networks (SNNs) present a bio-inspired, energy-efficient alternative. By communicating via sparse, binary "spikes," SNNs replace MACs with simpler accumulate (AC) operations. Despite their efficiency, SNNs have primarily been limited to classification tasks. Their application to complex regression problems, such as predicting precise displacement vectors for point cloud denoising, has remained largely unexplored. This paper introduces NI-SGCN, the first framework to leverage SNNs for 3D point cloud denoising, achieving significant energy savings without sacrificing competitive performance.
The NI-SGCN Framework
The authors propose two primary innovations to adapt SNNs for high-precision 3D denoising: Noise-Injected Integrate-and-Fire (NIIF) neurons and Noise-Injected Spiking Graph Convolutions (NI-SGC).
Noise-Injected Integrate-and-Fire (NIIF) Neurons
Standard SNN neurons typically operate deterministically. However, the authors discovered that injecting Gaussian noise into the neuron's membrane potential during training significantly enhances robustness. This stochasticity emulates biological brain functions and acts as a regularizer, helping the network manage the inherent "jitter" in noisy 3D data.
Noise-Injected Spiking Graph Convolution (NI-SGC)
Building on the EdgeConv operator, the NI-SGC method constructs local graphs around points to learn geometric relationships using sparse spiking signals. These are organized into Dense NI-SGC blocks to capture multi-scale features while maintaining low spike activity.
The framework supports two architectural variants:
- NI-PSGCN (Pure): A fully spiking network designed for maximum energy efficiency.
- NI-HSGCN (Hybrid): A hybrid model utilizing an SNN for feature extraction and a compact ANN for final "score estimation" (predicting noise gradients), striking a balance between accuracy and power consumption.
Experimental Setup and Key Insights
The models were evaluated on the PU-Net and PC-Net benchmarks, testing robustness across varying noise levels (0.5% to 2%) and point densities (10K to 50K points).
Several key insights emerged from the study:
- Stochastic Regularization: Noise injection prevents the SNN from over-sensitizing to raw scan perturbations.
- Information Preservation: The authors found that mean pooling is superior to max pooling in SNNs for this task. While max pooling on binary spikes often loses critical geometric detail, mean pooling converts these spikes into floating-point averages, preserving more information.
- Efficiency through Sparsity: The sparse nature of spikes allows hardware to remain inactive during periods of no signal, leading to substantial energy savings.
Results and Discussion
The proposed NI-SGCN framework demonstrates remarkable energy efficiency. The pure SNN variant (NI-PSGCN) is 12.75x more energy-efficient than the ANN-based ScoreDenoise model. The hybrid variant (NI-HSGCN) also offers a significant 4x efficiency improvement.
In terms of denoising quality, the hybrid model achieved a Chamfer Distance (CD) of 1.798 on the PU-Net dataset (2% noise), comparing favorably to the state-of-the-art ANN-based Pointfilter (1.175) while consuming far less power. Qualitative analysis confirms that the method effectively smooths noisy surfaces while maintaining sharp edges and intricate geometric details.
Conclusion
NI-SGCN marks a significant step forward in applying SNNs to complex 3D vision tasks. While a small accuracy gap remains between pure SNNs and top-tier ANNs for regression, the massive energy savings highlight the potential for neuromorphic hardware in real-world edge applications. Future research will focus on closing this performance gap to enable fully on-chip deployment on platforms like Intel’s Loihi.
Summary updated