Spatially-enhanced Spiking neural network for efficient point cloud analysis

Spiking Neural Networks (SNNs), with their spike-driven mechanism and low power consumption, attract extensive research attention in 2D visual tasks. For computationally intensive 3D point cloud tasks, SNNs exhibit greater potential in addressing the high computational complexity. However, SNNs hold significant application value and room for improvement. Unlike the 2D data with fixed spatial positions, complex and unordered points contain rich spatial information, posing significant challenges for spike feature modeling. We find that 3D spatial modeling is more crucial for SNN points analysis. Therefore, we analyze the key aspects of spiking spatial perception and introduce parameter-free Spiking Spatial Position Encoding (SSPE) to extract local positional information. Besides, the incorporation of Spiking Cross-feature Graph Position Encoding (SCGPE) is proposed to capture global spatial relationships. Our spatial enhancement is embedded within a framework centered on spiking fully connected layers, referred to as Spiking 3D Network (S3DNet). Extensive experiments demonstrate S3DNet's low energy consumption and state-of-the-art performance in SNNs. Specifically, with only 1.18M parameters, S3DNet achieves a classification accuracy of 92.34 % on the ModelNet40 and 84.49 % on the ScanObjectNN. Additionally, we explore SNN point cloud segmentation task for the first time and achieve an accuracy of 85.0 % on the ShapeNetPart with only 2.27M parameters. Overall, enhanced by spiking positional encoding, S3DNet has demonstrated the potential of Spiking Neural Networks (SNNs) in point cloud analysis.