Volume Rendering of Neural Implicit Surfaces

Yariv et al. introduce VolSDF, a volume-rendering framework in which the volume density is derived from a signed distance function rather than predicted by an independent MLP head as in NeRF (Mildenhall et al., 2020). Concretely, the density is the Laplace cumulative distribution function applied to the negative SDF, producing a thin shell of high density concentrated around the zero level set whose sharpness is controlled by a learnable bandwidth parameter.

The formulation has three consequences. First, the geometry inherits the inductive bias of an SDF — smooth, well-defined surfaces with usable normals — instead of NeRF's noisy density cloud. Second, the authors derive an explicit bound on the opacity-approximation error along a ray in terms of the bandwidth, enabling an importance sampler that places samples where they matter without redundant evaluations near empty space. Third, because shape (the SDF) and appearance (the radiance MLP) are now disentangled in the representation, scenes can be composed by swapping the geometry of one capture with the appearance of another.

On standard multi-view benchmarks VolSDF outperforms IDR (which requires masks) and matches or exceeds NeRF on view synthesis quality while producing reconstructions with substantially less geometric noise. The work builds on Implicit Geometric Regularization (Gropp et al., 2020), whose Eikonal loss is used to keep the SDF metric, and was published within days of NeuS (Wang et al., 2021), which solves the same problem with a different opacity construction.

VolSDF and NeuS are the foundational references for downstream SDF-from- multi-view methods in this corpus: MonoSDF, NeuralWarp, Voxurf, and the Gaussian-Splatting surface variants all inherit the Laplace-CDF density or one of its close cousins.