VGESP: Viewpoint-Induced Geometric Events for Spiking Point Cloud Understanding

Point cloud understanding has advanced rapidly, but the growing computational demands of modern 3D networks motivate more energy-efficient alternatives. Spiking neural networks (SNNs) provide a promising paradigm through sparse binary spikes and event-driven computation. However, existing spiking point cloud methods primarily focus on network-side feature processing, while the temporal structure of static point cloud inputs remains underexplored. To address this gap, we propose VGESP, a spike-driven point cloud framework that combines geometry-induced temporal representation with efficient global spiking attention. Specifically, we introduce Viewpoint-Induced Geometric Events (VGE), which convert visibility transitions and significant depth variations along short, ordered virtual-camera trajectories into sparse binary event streams, providing explicit geometric temporal cues for spiking computation. To efficiently integrate these temporal geometric cues, we further develop Vector Subtractive Spiking Attention (VSSA), which models geometric variations through vector-level discrepancies between binary spike features without explicitly constructing dense point-to-point attention maps. Extensive experiments demonstrate that VGESP achieves state-of-the-art performance among SNN-based point cloud methods on multiple benchmarks, outperforms many ANN-based counterparts, and retains competitive computational and energy efficiency.