DC-Net: Robust Point Cloud Normal Estimation with Dynamic Convolution

Normals are fundamental geometric attributes of three-dimensional (3D) objects, playing a crucial role in a variety of applications, including surface reconstruction and rendering. Existing methods often enhance network performance by increasing its depth and width, which leads to higher computational costs and presents challenges in accurately estimating normals, particularly in noisy regions. To address these limitations, this paper introduces DC-Net, a novel network specifically designed for robust and precise normal estimation in point clouds. Drawing inspiration from the Laplacian matrix used in graph convolutional networks, Unlike POAT-Net, which employs conventional self-attention, POAT-Net exhibits a discrepancy between its attention features and the original input features. Neglecting such discrepancy may cause the model to lose critical geometric details in complex curved surfaces. Different from DeepFit’s approach of pursuing accuracy through scale. we introduce an offset attention mechanism that refines feature representation by computing the offset between self-attentive features and input features. Additionally, dynamic convolution is incorporated to improve model performance without expanding the network’s depth or width. Finally, In contrast to AdaFit’s limitation of lacking global information fusion we present a Global Bilinear Fusion module, which preserves global information and enhances the network’s resilience to noise by computing a bilinear combination of global channels and point descriptors. Experimental results demonstrate that our method consistently outperforms the compared approaches on both synthetic and real-world datasets. Notably, the average angular root mean squared error (RMSE) on the PCPNet and SceneNN datasets are 10.02 and 9.72, respectively, with notable improvements observed on noisy point cloud data.