Image-based modeling of inhomogeneous single-scattering participating media

Abstract

We propose a method to reconstruct inhomogeneous single-scattering participating media, which could preserve fine high-frequency details of density field. Volumetric data and the ratio of absorption coefficient to scattering coefficient are used to describe the spatial-varying density distribution and optical properties of certain participating media, and a function between the above parameters and captured pixel values is built. Thus the problem of how to find solutions to these parameters is formulated into a nonlinear numerical optimization problem. In order to reduce large time overheads and numerical instability brought by simultaneously solving large numbers of voxels, we propose an initialization algorithm for enabling the assigned density values to satisfy the regularity of brightness distribution in captured images approximately as well as a progressive refinement algorithm for multi-resolution volumetric data. Besides, we propose a parallel multi-voxel gradient computation algorithm of utilizing hardware acceleration to reduce time overheads in gradient computation for large numbers of voxels. Experiment results indicate that our method is well suited for reconstructing thin smoke, retaining high-frequency details from images captured from multiple viewpoints.