Mollifying Realistic Image Synthesis for Time Constrained Rendering
Final Report Abstract
Realistic image synthesis, or rendering, is a key problem in computer graphics with numerous application areas, from digital entertainment to architecture and industrial design. The goal of most rendering algorithms is to converge to the ground truth solution as quickly as possible with increasing computation time. In light transport simulation nowadays Monte Carlo-integration is used to compute a solution to the radiative transport. The computation typically continues until desired quality criteria are met, but the desired computation time is often not an input to the renderer. In contrast to this, we are considering the problem to obtain the best possible approximate solution when the computation time is limited upfront. The key idea is to achieve this by modifying the rendering problem at hand, that is the scene with its geometries and materials and the light transport simulation, to obtain an easier problem to be solved. This is often called “mollification”. Solving the mollified problem may in general not converge to the ground truth solution and a systematic error (“bias”) remains, but it will provide a better approximation with lower overall error given a limited computation time, compared to solving the original problem. In the course of the project, mollification has been applied to several aspects relevant to modern image synthesis algorithms. First, we addressed the problem of rendering images from scenes with challenging geometry and material configurations. We developed means to separate the light transport whilst computing it into components which are easy, and those which are difficult to compute. A mollified representation of the latter implicitly captures the entire scene configuration and enables us to sample, and thus compute, the difficult part of the simulation much more robustly and efficiently. The second focus was the mollification of spectral light transport simulation with fluorescent materials. Such simulations and materials are of utmost importance to render scenes with highest possible realism regarding light-matter interactions and illumination. For this, we developed and evaluated new material models, means to efficiently create plausible spectral data from typical RGB-input, and improved bidirectional light transport simulation by mollifying the path connections required to sample the space of possible light transport paths. The impact of these mollifications has been evaluated to assess the trade-off between systematic bias and convergence speed. In addition, we developed a method to help detect systematic errors in Monte Carlo-image synthesis methods.
Publications
- A Simple Diffuse Fluorescent BBRRDF Model, in Workshop on Material Appearance Modeling, R. Klein und H. Rushmeier, Hrsg., The Eurographics Association, 2018, ISBN: 978-3-03868-055-0
A. Jung, J. Hanika, S. Marschner und C. Dachsbacher
(See online at https://doi.org/10.2312/mam.20181193) - Selective Guided Sampling with Complete Light Transport Paths, Transactions on Graphics (Proceedings of SIGGRAPH Asia), Jg. 37, Nr. 6, Dez. 2018
F. Simon, A. Jung, J. Hanika und C. Dachsbacher
(See online at https://doi.org/10.1145/3272127.3275030) - Wide Gamut Spectral Upsampling with Fluorescence, Computer Graphics Forum, 2019, ISSN: 1467-8659
A. Jung, A. Wilkie, J. Hanika, W. Jakob und C. Dachsbacher
(See online at https://doi.org/10.1111/cgf.13773) - Detecting Bias in Monte Carlo Renderers using Welch’s t-test, Journal of Computer Graphics Techniques (JCGT), Jg. 9, Nr. 2, S. 1–25, Juni 2020
A. Jung, J. Hanika und C. Dachsbacher
(See online at https://doi.org/10.5445/ir/1000120188) - Improving Spectral Upsampling with Fluorescence, in Workshop on Material Appearance Modeling, R. Klein und H. Rushmeier, Hrsg., The Eurographics Association, 2020, ISBN: 978-3-03868-108-3
L. König, A. Jung und C. Dachsbacher
(See online at https://doi.org/10.2312/mam.20201139) - Spectral Mollification for Bidirectional Fluorescence, Computer Graphics Forum, 2020, ISSN: 1467-8659
A. Jung, J. Hanika und C. Dachsbacher
(See online at https://doi.org/10.1111/cgf.13937)
