NeUDF: Learning Neural Unsigned Distance Fields With Volume Rendering.

Area of Science:

• Computer Vision and Graphics
• Geometric Deep Learning
• 3D Reconstruction

Background:

• Neural implicit rendering has advanced multi-view shape reconstruction.
• Existing methods using signed distance functions (SDFs) are limited to closed surfaces.
• Real-world objects often possess open-surface structures, posing a challenge for current methods.

Purpose of the Study:

• To introduce a novel neural rendering framework, NeUDF, capable of reconstructing surfaces with arbitrary topologies.
• To enable reconstruction solely from multi-view supervision, overcoming limitations of SDF-based methods.
• To address the challenge of reconstructing open surfaces, which are common in real-world objects.

Main Methods:

• NeUDF utilizes the unsigned distance function (UDF) for flexible representation of arbitrary surfaces.
• Formalized rules for neural volume rendering adapted for open surface reconstruction (self-consistent, unbiased, occlusion-aware).
• Developed a dedicated rendering weight function tailored for UDF and a normal regularization strategy to resolve orientation ambiguity in open surfaces.

Main Results:

• NeUDF successfully reconstructs surfaces with arbitrary topologies from multi-view data.
• The method significantly outperforms state-of-the-art techniques on challenging datasets, particularly those with open boundaries.
• Evaluated on MGN and Deep Fashion 3D datasets, demonstrating superior performance in complex shape reconstruction.

Conclusions:

• NeUDF provides a robust solution for multi-view shape reconstruction of objects with open surfaces.
• The UDF representation and tailored rendering techniques overcome limitations of previous SDF-based approaches.
• This work expands the applicability of neural rendering to a wider range of real-world 3D objects.