Unveiling the Power of Dual Positional Encoding in Point Cloud Learning with DPAtteNet

Traditional attention mechanisms in point cloud learning typically use a single type of positional encoding, which captures positional relationships within individual neighborhoods but neglects the spatial relationships between neighborhoods. Understanding these spatial relationships is crucial for grasping the overall shape of the point cloud and improving feature extraction. To address this, we propose Dual Positional Encoding Attention (DPAttention). DPAttention combines global positional encoding to capture spatial relationships between neighborhoods with local positional encoding to address positional relationships within neighborhoods. This approach enables the model to effectively understand both the global geometric structure and the local details of the point cloud, thus enriching feature extraction through the attention mechanism. Building on DPAttention, we introduce DPAtteNet, a deep learning network designed for point cloud classification and segmentation. We evaluated DPAtteNet on the ModelNet40, ScanObjectNN, and ShapeNet Part datasets. The experimental results confirm the effectiveness of our method. The code will be available at https://github.com/NWUzhouwei/DPAtteNet.