Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields

Barron et al. extend Mip-NeRF (Barron et al., 2021) from object-centric capture to unbounded 360-degree scenes — outdoor environments and large indoor spaces where the camera rotates around a point and the scene content extends to arbitrary depths. Three problems prevent vanilla Mip-NeRF from handling this setting: distant background content cannot be parameterized by a uniformly sampled bounded volume; the optimization is slow on large scenes; and unbounded captures admit floaters and shape-radiance ambiguities that produce visually disturbing artifacts.

The paper introduces three contributions. A non-linear scene parameterization contracts unbounded 3D coordinates into a bounded domain, with a piecewise rule that keeps the near-field nearly linear and compresses the far-field smoothly toward infinity — extending the classical NDC trick to non-forward-facing captures. An online distillation scheme uses a small proposal MLP to predict where mass lies along a ray, supervised by a larger radiance MLP, which makes hierarchical sampling far more efficient and lets the radiance network focus capacity on visible content. A distortion-based regularizer penalizes radiance contributions that are smeared along the ray rather than concentrated, which suppresses the floaters and background haze that plague large-scene NeRFs.

The combination reduces mean-squared error by 57% relative to Mip-NeRF on the new 360-degree benchmark that the authors introduce with the paper, and produces clean depth maps for highly intricate real-world captures. The "mipnerf360" dataset itself has become a standard evaluation suite for unbounded radiance-field methods, used by every subsequent system in this corpus that targets the same regime including 3D Gaussian Splatting, MERF, and Zip-NeRF.

The non-linear contraction and proposal-network architecture are now ubiquitous components of any modern radiance-field codebase.