Surf-NeRF: Surface Regularised Neural Radiance Fields
This work improves geometric accuracy for NeRF-based scene representations, which is incremental but important for geometry-critical applications like 3D reconstruction.
The paper tackles the problem of shape-radiance ambiguity in Neural Radiance Fields (NeRFs) by introducing surface regularization techniques, resulting in 28% more accurate normals than traditional grid-based NeRF variants.
Neural Radiance Fields (NeRFs) provide a high fidelity, continuous scene representation that can realistically represent complex behaviour of light. Despite works like Ref-NeRF improving geometry through physics-inspired models, the ability for a NeRF to overcome shape-radiance ambiguity and converge to a representation consistent with real geometry remains limited. We demonstrate how both curriculum learning of a surface light field model and using a lattice-based hash encoding helps a NeRF converge towards a more geometrically accurate scene representation. We introduce four regularisation terms to impose geometric smoothness, consistency of normals, and a separation of Lambertian and specular appearance at geometry in the scene, conforming to physical models. Our approach yields 28% more accurate normals than traditional grid-based NeRF variants with reflection parameterisation. Our approach more accurately separates view-dependent appearance, conditioning a NeRF to have a geometric representation consistent with the captured scene. We demonstrate compatibility of our method with existing NeRF variants, as a key step in enabling radiance-based representations for geometry critical applications.