Multi-grid representation with field regularization for self-supervised surface reconstruction from point clouds

Abstract

Surface reconstruction from point clouds plays a crucial role in computer vision. The current state-of-the-art methods solve this problem by learning signed distance functions (SDFs) with ground truth distance supervisions, which are difficult to obtain. Moreover, most recent works represent each shape with a single or several latent codes, which fail to provide detailed guidance to reconstruct the local geometry. To address these issues, we propose MGSDF, a novel method for high-fidelity and fast surface reconstruction from raw point clouds. Specifically, we design a scalable representation with learnable hierarchical feature grids to capture multi-level geometric details. We introduce a self-supervised learning scheme that optimizes the SDF directly from the raw point cloud by pulling the space onto the surface. In addition, we propose a field regularization constraint on the predicted distance values and gradients on the zero-level set of SDFs for robust optimization. Our experimental results demonstrate significant improvements over the state-of-the-art in surface reconstruction from clean, noisy and varying density point clouds under widely used benchmarks.