PARS-Path recycling and sorting for efficient cloud tomography

Abstract

Inverse rendering estimates scene characteristics from image data. We derive an efficient framework for inverse rendering and specifically computed tomography (CT) of volumetric scattering objects. We focus on clouds, which have a key role in the climate system and require efficient analysis at a huge scale. Data for such reconstruction are multiview images of each cloud taken simultaneously. This acquisition mode is expected by upcoming future spaceborne imagers, such as CloudCT. Prior art shows that scattering CT can rely on Monte–Carlo (MC) light transport. This approach usually iterates differentiable radiative transfer, requiring many sampled paths per iteration. We present an acceleration approach: path recycling and sorting (PARS). It efficiently uses paths from previous iterations for estimating a loss gradient at the current iteration. This reduces the iteration run time. PARS enables further efficient realizations. Specifically, sorting paths according to their size accelerates implementations on a graphical processing unit (GPU). PARS, however, requires a correction operation for unbiased gradient estimation. This can be achieved by utilizing a well-established concept from MC integration methods, as we show in this paper. We derive the theory of PARS and demonstrate its efficiency on cloud tomography of both synthetic and real-world scenes. Moreover, we demonstrate PARS on simple reflectometry examples.