Surface Reconstruction from Raw Point Cloud via Energy-Based Models

Abstract

Implicit representations via neural networks are emerging as a method for representing 3D shapes due to their memory efficiency and representational capability. Thus, methods for surface reconstruction by learning or generating implicit representations from 3D point cloud data are being actively studied. In this work, we propose a method to learn implicit surface representations using energy-based models (EBMs). We consider the point cloud sampled from a three-dimensional continuous distribution modeled by the EBM and optimize the energy function by likelihood maximization. In this optimization process, using a suitable energy function based on neural networks, the implicit representation of the target shape can be easily derived from the optimized energy function. Additionally, appropriate parameter settings of the EBM can make the model more robust to the noise of point clouds. Our experiments confirmed that the method is robust to point cloud noise and almost comparable to conventional surface reconstruction methods.