Implicit Geometric Regularization for Learning Shapes

Gropp et al. propose Implicit Geometric Regularization (IGR), a training objective for fitting neural signed distance functions directly to raw point clouds without supervised ground-truth occupancy. The loss has two terms: a zero-set term encouraging the network to vanish on input points (optionally with surface normals), and an Eikonal regularizer encouraging the gradient of the network to have unit norm everywhere in the ambient space.

The Eikonal term is the methodological core. Because the loss is non-zero almost everywhere in space, the optimizer cannot collapse to the degenerate solution where the network is zero outside the support of the point cloud. The authors prove, in the linear regime, that the regularizer biases the network toward smooth zero level sets — the kind of clean, manifold surfaces practitioners want — rather than the noisy, fragmented implicit surfaces produced by direct chamfer-style fitting.

Empirically, IGR recovers high-fidelity surfaces from sparse oriented points and unoriented points alike, outperforming DeepSDF-style methods that require pre-computed ground-truth SDF samples and Occupancy Networks that need explicit inside/outside labels. The method became a building block for IDR, NeuS, VolSDF, and MonoSDF: any neural-implicit-surface method that backs a radiance field with an SDF inherits IGR's Eikonal regularizer to keep the distance field well-behaved.

The contribution is a clear example of how a carefully chosen regularization term, applied to coordinate-based networks, can replace expensive pre-processing pipelines that produce explicit supervision. Within the 3D reconstruction corpus this paper is the bridge between classical level-set methods and the volume-rendering era surface reconstruction lineage led by NeuS and VolSDF.