NeuralUDF

NeuralUDF

Learning Unsigned Distance Fields for Multi-view Reconstruction of Surfaces with Arbitrary Topologies

Background

  • SDF and OF (occupancy) : implicit

  • Marching Cube: extract mesh from implicit and expensive

Motivation

  • SDF 局限于 closed surfaces

Overview

  • UDF and density function of the volume rendering
  • Differentiable indicator function
  • iso-surface regularizer

Unbiased volume density function

  • SDF 20231009-NeuralUDF-2023-10-09-16-51-53
  • UDF 20231009-NeuralUDF-2023-10-09-16-53-59 \(\Phi(t)\) is the indicator.

Visibility indicator function

Use soft probabilistic distribution instead of using hard \(\epsilon\) bound

20231009-NeuralUDF-2023-10-09-17-03-21 20231009-NeuralUDF-2023-10-09-17-03-57

Gradient-aware indicator

增加 UDF 梯度 Mask 50%

bg right 90%

Iso-surface regularizer

20231009-NeuralUDF-2023-10-09-17-20-54 \(M\) is the number of samples. 产生更清晰的等值面。

overall loss function

20231009-NeuralUDF-2023-10-09-17-26-17