Surface normal estimation is a basic task for many point cloud processing algorithms. However, it can be challenging to capture the local geometry of the data, especially in presence of noise. Recently, deep learning approaches have shown promising results. Nevertheless, applying convolutional neural networks to point clouds is not straightforward, due to the irregular positioning of the points. In this paper, we propose a normal estimation method based on graph-convolutional neural networks to deal with such irregular point cloud domain. The graph-convolutional layers build hierarchies of localized features to solve the estimation problem. We show state-of-the-art performance and robust results even in presence of noise.
