Local Regression Mixture Models for Compact and Adaptive Plenoptic Representation

Abstract

We propose local regression mixture models as an adaptive and compact continuous radiance field representation. The whole extent of the scene is covered densely by multivariate 3D Gaussian components and associated linear regressors encode the means and local gradients of the radiometric responses. The partitions of unity from these marginal kernels act as gating functions that mix local regressors into a continuous volume that is explicitly defined everywhere. Both sharp or smooth regions may be captured by the relative overlap between density kernels. Our preliminary experiment on synthetic models shows that it is possible to represent manifolds with far fewer parameters than classic or splatting methods. We compare the result of raytraced numerical integration using two variants for the model representation: volumetric density (rho) or volumetric opacity (alpha).