Abstract

Scale-ambiguity in 3D scene dimensions leads to magnitude-ambiguity ofvolumetric densities in neural radiance fields, i.e., the densities double whenscene size is halved, and vice versa. We call this property alpha invariance.For NeRFs to better maintain alpha invariance, we recommend 1) parameterizingboth distance and volume densities in log space, and 2) adiscretization-agnostic initialization strategy to guarantee high raytransmittance. We revisit a few popular radiance field models and find thatthese systems use various heuristics to deal with issues arising from scenescaling. We test their behaviors and show our recipe to be more robust.

 

Quick Read (beta)