Differentiable point process texture basis functions for inverse procedural modeling of cellular stochastic structures

Abstract

To meet the growing demand of rich 3D content, the Computer Graphics industry needs tools to automatically create procedural textures and materials from image exemplars. In this paper, we focus on the inverse procedural modeling of cellular stochastic structures, i.e., spatial distributions of repetitive, possibly blending similar shapes over a planar surface, resulting from stochastic processes. Such structures are very frequent in natural textures and materials, that exhibit random spatial variations. We represent cellular stochastic structures procedurally using thresholded Point Process Texture Basis Functions (PPTBFs). Previous approaches that learn PPTBF representations of structures maps solely rely on sampled data obtained by uniformly sampling the PPTBF parameter space, and parameter prediction based on these representations fails when the exemplars are visually too far from the training datasets. For the specific class of cellular stochastic structures, we propose to overcome this limitation by introducing a Cellular PPTBF, or C-PPTBF, defined with differentiable window and feature functions. Based on this procedural model, we present a fully differentiable pipeline and optimization procedure to automatically estimate the parameters of a C-PPTBF. We show that our method is efficient (between 2 and 5 min per exemplar) and yields more robust parameter prediction than the state-of-the-art DiffProxy approach.