Daily digest — 2026-04-20

Surfaces day. Radiance fields are excellent at view synthesis and bad at giving you a mesh; the surface-reconstruction line is the field's answer.

Implicit Geometric Regularization (Gropp et al., 2020) is the unsexy prerequisite: a loss that encourages a learned implicit function to behave like a signed distance function, via an Eikonal term on its gradients. Almost every later SDF-based paper in this corpus cites it as the regulariser that made unconstrained shape learning stable.

The 2021 surface trilogy is where it gets interesting. UNISURF (Oechsle et al.) was first out: unify implicit surfaces and radiance fields by replacing density with an occupancy function. Then NeuS (Wang et al.) and VolSDF (Yariv et al.) — published days apart — both argued that occupancy biases the surface estimate and proposed unbiased SDF-to-density conversions. NeuS's derivation is the cleaner read. The Totoro97/NeuS implementation has become the de-facto baseline in the surface-reconstruction benchmarks.

MonoSDF (Yu et al., 2022) is the practical follow-on: feed an off-the-shelf monocular depth and normal predictor into the surface optimiser as additional cues, so the SDF training converges with far fewer views and on indoor scenes where the purely-multi-view methods struggle.

The detour was BARF (Lin et al., 2021) on bundle-adjusting NeRF — optimising camera poses jointly with the field via a coarse-to-fine positional encoding schedule. Not strictly a surface paper, but it answers the "what if SfM poses are bad?" question that every surface method ducks. Same trick reappears in the SLAM literature later.

Six papers, ~50 min. The surface line and the splatting line are about to collide; that's next week's reading.