DiffSDF: Learning Implicit Surface from Noisy Point Clouds

Computing surface properties such as surface normals, curvatures, and reconstructing surfaces from noisy point clouds are a few key tasks in 3D point cloud processing. UUsing theSigned Distance Function (SDF) to represent the implicit surface of a point cloud is an efficient way to accomplish these tasks at the same time. Recently, it has been shown that SDF can be effectively learned by an MLP neural network. However, existing works use the distance values as the supervised signal only, which cannot guarantee the learned to be a valid sign distance function. Some methods assume all the input points are on the zero level, hence they cannot deal with noisy point clouds. In this paper, we propose a novel approach, namely DiffSDF, which takes a noisy point cloud as input and estimates a continuous SDF that implicitly represents the underlying shape for the input data. In addition, we supervise the gradient of the SDF, calculated from differentiating the MLP network with respect to the input. Through experiments, we show that our approach achieves superior results in multiple raw point cloud processing tasks.