UNISURF: Unifying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction

Oechsle et al. propose UNISURF, the first method to unify implicit-surface reconstruction (DVR, IDR) and radiance-field volume rendering (NeRF) under a single formulation. The motivating tension is that DVR/IDR-style surface methods reconstruct clean geometry but require accurate per-pixel object masks as supervision, while NeRF requires no masks but produces a fuzzy volume density that does not admit a clean surface.

The unifying observation is that surface rendering and volume rendering differ only in how opacity is distributed along a ray. UNISURF parameterizes the scene as an occupancy field (binary inside/outside, relaxed to a probability) and lets a single training pipeline interpolate between the two regimes: early in training the model behaves like a soft volume renderer, so gradients flow through entire rays and bad initializations recover; late in training the opacity sharpens around the occupancy boundary and the renderer collapses to surface rendering, yielding crisp, mask-free reconstructions.

Experiments on DTU, BlendedMVS, and an internal synthetic indoor dataset show that UNISURF outperforms NeRF on reconstruction quality while matching IDR (Yariv et al., 2020) without using foreground masks. The unification also enables more efficient ray sampling: once the model has learned where the surface lies, the sampler can concentrate points around the occupancy boundary rather than uniformly along the ray.

UNISURF is a foundational reference for the surface-from-volume lineage that includes NeuS (Wang et al., 2021) and VolSDF (Yariv et al., 2021), both of which substitute a signed distance function for UNISURF's occupancy field and refine the bias analysis. Together these three papers established the multi-view surface-reconstruction paradigm that dominates the area today.