Neural Surface Detection for Unsigned Distance Fields

TL;DR

This work introduces a deep-learning approach to taking a UDF and turning it locally into an SDF, so that it can be effectively triangulated using existing algorithms, and shows that it achieves better accuracy in surface detection than existing methods.

Abstract

Extracting surfaces from Signed Distance Fields (SDFs) can be accomplished using traditional algorithms, such as Marching Cubes. However, since they rely on sign flips across the surface, these algorithms cannot be used directly on Unsigned Distance Fields (UDFs). In this work, we introduce a deep-learning approach to taking a UDF and turning it locally into an SDF, so that it can be effectively triangulated using existing algorithms. We show that it achieves better accuracy in surface detection than existing methods. Furthermore it generalizes well to unseen shapes and datasets, while being parallelizable. We also demonstrate the flexibily of the method by using it in conjunction with DualMeshUDF, a state of the art dual meshing method that can operate on UDFs, improving its results and removing the need to tune its parameters.

Figures (16)

• Figure 1: Neural Surface Detection. We mesh a UDF by turning it into a pseudo-SDF using a neural network, and then meshing it using a triangulation algorithm. When coupled with Marching Cubes Lewiner03, our method yields better results than other MC-based methods. When coupled with DualMeshUDF Zhang23b, a dual method, it removes the requirement for parameter-tuning and closes most holes in the reconstructions.
• Figure 2: Neural Surface Detection. We formulate the surface detection problem as a per-cell classification task. In each cell, we map point distances and gradients to a sign configuration of the cell vertices, which can be used to mesh the surface via Marching Cubes Lewiner03 or Dual Contouring Ju02.
• Figure 3: Reconstruction methods based on Marching Cubes. We use our method to transform UDFs into pseudo-SDFs, which we mesh using MC. We compare our results to MC-based baselines on the ABC, MGN, and ShapeNet Cars datasets at various resolutions. More reconstructions are provided in the supplementary material.
• Figure 4: Reconstruction methods based on Dual Contouring. We use our method in conjunction with DualMeshUDF. We compare our results to Unsigned Neural Dual Contouring, DualMeshUDF with default parameters, and DualMeshUDF with tuned parameters.
• Figure 5: Reconstructing thin surfaces at very low resolutions. When trained with noise augmentation, our approach can reconstruct thin surfaces, here at resolution 32, which are missed by Marching Cubes.