3DSMT++: Towards High-Performance Spike-Driven Point Cloud Learning

Spiking Neural Networks (SNNs) offer a promising energy-efficient perception paradigm via sparse, event-driven computation. Recently, deep SNNs have made strong progress in 2D visual representation learning. However, extending these advances to sparse and irregular 3D point clouds remains challenging. Existing spiking point cloud models still lag behind strong Artificial Neural Network (ANN) methods. This is largely because sparse spike representations fail to fully capture rich geometric structures in point clouds. To address this issue, we propose 3DSMT++, an enhanced hybrid Spiking Mamba-Transformer for high-performance and energy-efficient point cloud analysis. First, we introduce Multi-Scale Spiking Patch Embedding (MSPE) to extract complementary geometric features at different spatial scales, together with masked pre-training to improve multi-scale representation learning. We further propose Spiking Distance-aware Offset Attention (SDOA), which leverages spatial relations and feature variations of neighboring points for precise local geometric modeling. In addition, we design a Spiking Point Mamba (SPM) architecture to improve efficient global feature interaction for unordered point clouds. Overall, 3DSMT++ improves multi-scale, local, and global representations while retaining event-driven computational efficiency. Extensive experiments on point cloud benchmarks show that our method achieves SOTA performance among SNN point cloud methods, outperforms many ANN approaches, and maintains competitive computational and energy efficiency.