A dual-stream edge-aware network with collect-and-disperse attention for 3D point cloud classification

Abstract

3D point cloud data from optical sensing systems suffers from inherent irregularity and sparsity, which impairs the effective fusion of fine-grained local details and global contextual features in deep learning models. To address this issue, we propose a fully-supervised 3D point cloud classification framework integrating an edge-aware sampling strategy with a multi-scale attention mechanism. First, edge-aware downsampling preserves key contour and structural information by analyzing local geometric variance, mitigating information loss in non-uniform point distributions. A dual-stream architecture is then designed for multi-scale feature representation: the global module adopts adaptive graph convolution and a hierarchical Collect-and-Disperse mechanism to capture long-range dependencies and realize global local feature coupling, while the local module leverages an offset-attention structure with relative positional encoding store fine geometric details. Extensive experiments on the cluttered ScanObjectNN and standard ModelNet40 datasets demonstrate the superiority of our method, achieving classification accuracies of 91.2% and 94.6% respectively, outperforming state-of-the-art approaches. This framework enhances feature discriminability and robustness for point cloud classification, providing a robust algorithmic foundation for intelligent optical sensing systems in practical 3Dperception tasks.