Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces

Baorui Ma, Zhizhong Han, Yu-Shen Liu, Matthias Zwicker

Introduction

Reconstructing continuous surfaces from discrete 3D point clouds is a fundamental challenge in geometry processing. Signed Distance Functions (SDFs) have emerged as a powerful representation, encoding a shape by the distance from any point in space to the nearest surface, with a sign indicating whether the point is "inside" or "outside."

However, training neural networks to learn SDFs typically requires ground truth distance values, which are difficult to obtain for raw, unoriented point clouds. Previous methods often defaulted to learning unsigned distance fields, which fail to distinguish between interior and exterior, or relied on complex geometric constraints. Neural-Pull introduces a simple yet effective objective to learn high-quality, consistent SDFs directly from raw points.

Methodology: The Pulling Operation

The core innovation of Neural-Pull is the pulling operation. Rather than directly regressing a distance value, the framework trains a neural network $f$ to move (or "pull") any 3D query location $q$ to its closest point on the surface using the network's own predicted distances and gradients.

The pulling operation: Query points $q_1$ and $q_2$ are moved to surface points $t_1$ and $t_2$ using the network's predicted distance $s$ and gradient $\nabla f$.

For a query point $q$, the network predicts a signed distance $s$ and computes the gradient $\nabla f$ via backpropagation. The point is then updated to a new location $t'$: $$t' = q - f(q) \times \frac{\nabla f(q)}{\|\nabla f(q)\|_2}$$

The network is optimized by minimizing the squared distance between the "pulled" point $t'$ and its actual nearest neighbor in the input point cloud. This self-consistent loop forces the network to align its internal distance and gradient logic with the physical geometry of the surface.

Theoretical Insights: Gradient Continuity

A critical distinction between signed and unsigned distance fields lies in the behavior of their gradients. In a signed field, the gradient direction remains continuous as one crosses the surface. In contrast, an unsigned field's gradient flips direction at the surface, creating a discontinuity.

Comparison of Signed (a, b) and Unsigned (c, d) distance fields. The signed field maintains a continuous gradient (b), whereas the unsigned field exhibits a jump (d).

Neural-Pull leverages this property by requiring the gradient to point consistently to successfully pull points from both sides of the surface. This objective naturally encourages the network to converge to a valid SDF without requiring explicit Eikonal loss terms or other complex regularizations.

Experimental Results

Neural-Pull was evaluated on benchmarks including the ABC, FAMOUS, and ShapeNet datasets. It significantly outperformed existing methods like IGR and Points2Surf in capturing fine details and thin structures. For instance, on the FAMOUS dataset, Neural-Pull reduced the mean L2-Chamfer Distance from 1.08 (IGR) to 0.35.

Beyond surface reconstruction, the model demonstrates strong performance in conditional tasks, such as single-image 3D reconstruction, where it successfully recovers complex topologies from 2D inputs.

Conclusion and Future Work

Neural-Pull provides a robust framework for learning implicit surfaces from raw data. While highly effective, the method's reliance on nearest-neighbor searches during training can be computationally intensive for large-scale point clouds. Future research may focus on optimizing these searches or developing more sophisticated query sampling strategies to enhance efficiency and robustness against extreme noise.

Summary updated

BibTeX
@inproceedings{ma2021neural,
  author = {Baorui Ma and Zhizhong Han and Yu-Shen Liu and Matthias Zwicker},
  title = {{Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces}},
  booktitle = {International Conference on Machine Learning},
  year = {2021},
  url = {https://openreview.net/forum?id=ZXzTx1XxYOT},
}
Figure