Reproduce “Neural Unsigned Distance Fields for Implicit Function Learning” on OpenResearch

About this paper

Julian Chibane, Aymen Mir, Gerard Pons-Moll

Neural Unsigned Distance Fields (NDF) introduces a neural implicit representation that learns the unsigned distance to a surface, overcoming the limitation of prior methods to only represent closed surfaces. The approach achieves state-of-the-art 3D reconstruction performance on benchmark datasets while also demonstrating successful capture of open surfaces, complex internal geometries, and general mathematical functions.

Key insights

• By learning Unsigned Distance Fields (UDFs) instead of Signed Distance Functions (SDFs) or occupancy, NDF can represent a far broader class of shapes, including open surfaces and those with complex internal structures, without requiring artificial data closures.
• The analytical computation of spatial gradients from the learned UDF enables robust iterative projection for extracting dense point clouds and provides a direction towards the surface.
• Modified sphere tracing, traditionally a graphics technique, can be repurposed as a machine learning tool for direct rendering and even multi-target regression for complex functions and manifolds.

Summary by alphaXiv.