Traditional 3D representations often force a choice between mathematical elegance and topological flexibility. While Signed Distance Functions (SDFs) are the gold standard for representing solid, "watertight" volumes, they fail when faced with open surfaces like a single leaf, a piece of clothing, or the complex interior of a car. To address this, researchers have pivoted toward Unsigned Distance Fields (UDFs), a more general representation that can describe surfaces of arbitrary topology.
The Problem of the Non-Differentiable Kink
The core intuition behind a UDF is simple: for any point in space, the field stores the distance to the nearest point on the surface. Because distances are always non-negative, the function reaches a minimum of zero exactly at the surface interface.
However, this creates a significant mathematical hurdle for neural networks. As a point crosses the surface, the distance decreases to zero and then immediately begins to increase again. This forms a sharp "cusp" or "kink" in the function. Mathematically, the UDF is non-differentiable at the exact location we care about most: the zero level set.
Rahul Venkatesh and Sarthak Sharma note in Deep Implicit Surface Point Prediction Networks that this lack of differentiability causes gradients to vanish or flip discontinuously. For a neural network optimized via gradient descent, this "shatters" the surface, leading to the fragmented, noisy, or "choppy" reconstructions seen in early UDF models.
Moving from Scalars to Vectors
To overcome the instability of the zero level set, recent work shifts the focus from predicting a single scalar distance to predicting geometric vectors.
- Closest Surface Points (CSP): Instead of just asking "how far is the surface?", Venkatesh et al. propose predicting the exact coordinates of the nearest point. This provides a much richer signal for the network to learn.
- Gradient Distance Functions (GDF): Federico Stella and Hieu Giang Le introduced the Gradient Distance Function, which predicts a vector pointing toward the surface. Unlike the scalar distance, this vector field remains differentiable at the surface, allowing for smoother optimization and faster training.
- Neural Shortest Path (NSP): Yesom Park and colleagues further this by learning the Neural Shortest Path, a vector-valued representation that simultaneously optimizes for distance and direction, providing robustness even with sparse or noisy point cloud data.
Smoothing the Boundary with Projection and Regularization
Another strategy involves using the reliable information near the surface to fix the unreliable information at the surface.
In Learning a More Continuous Zero Level Set in Unsigned Distance Fields through Level Set Projection, Baorui Ma and Zhizhong Han introduce Level Set Projection. This technique uses the smoother gradients from points slightly away from the surface to "pull" the zero level set into a continuous manifold, preventing the "shattered" artifacts common in standard UDFs.
Similarly, Charles C. Yang and Chuanxiang Yang utilize Monge-Ampere Regularization. By applying this partial differential equation as a constraint, they force the neural field to maintain mathematical regularity, overcoming the non-differentiability of the distance function and enabling high-fidelity reconstruction of both watertight and open shapes.
Reconstructing Reality from Images
The ultimate goal of these representations is often multi-view reconstruction. NeuralUDF, developed by Xiaoxiao Long and Cheng Lin, adapts volume rendering techniques to work with UDFs. By introducing a gradient-aware rendering scheme, they can reconstruct complex topologies—like the thin, delicate structure of a leaf—directly from 2D images, a feat previously impossible with standard SDF-based neural renderers.