Light Transport Simulation with Non-Monte Carlo Approaches

当今数字媒体行业中的内容创建在很大程度上依赖于真实感图像的创建。示例应用包括电影、工业设计和科学可视化。通过模拟光在真实的世界中的传播（通常称为光传输模拟）来生成这种真实感图像。除了数字媒体之外，光传输模拟也正在成为分析灯光观测数据（传感器测量）和通过生成逼真的虚拟街景来训练自动驾驶车辆的主要工具。 光传输模拟计算到达传感器（或眼睛）的光的强度，给定光源的输入数据和物体的外观和形状。由于光可以从物体上反射任意次数，因此从光源发出的光到达传感器的可能路径有无限多条。光传输模拟通常通过随机选择若干路径来近似这样的无限路径集合，并且然后使用那些随机选择的路径来估计到达传感器的光的强度。这种方法通常被称为蒙特卡罗方法。虽然在数字媒体行业中广泛使用，但已知使用这种蒙特卡罗方法的光传输模拟具有几个基本问题。最常见的问题是，由于其数值误差，它会产生嘈杂的图像。由于照明通常是平滑的，因此噪声是应当被去除的不期望的伪影。这也与蒙特卡罗方法很少假设光的可能路径这一事实有关，因为光的可能路径确实有很多限制。例如，光的能量和传感器处的强度都是非负的，但蒙特卡罗并没有明确使用这一事实。蒙特卡罗方法也忽略了光传输的递归性质。例如，三次反弹后的照明应该与四次反弹后的照明相似，但Monte Carlo方法独立地模拟不同数量的反弹。这项研究计划的目的是挑战普遍认为蒙特卡罗是光传输模拟的唯一可行的方法。我们的目标是通过开发新的方法，如果我们只依赖于传统的蒙特卡罗方法是不可能的，以解决应用蒙特卡罗光输运模拟的基本问题。我和我的学生通过探索各种不适合传统蒙特卡罗范式的方法来并行研究不同的问题。这项研究计划开辟了一个新的模式，光传输模拟的基础上非蒙特卡罗方法。结果有助于对光输运模拟的基本理解，并导致更有效的光输运模拟。这种高效的光传输模拟是必要的，以适应数字媒体行业对视觉细节不断增长的需求。

Content creation in the digital media industry today relies heavily on the creation of photorealistic images. Example applications include movies, industry design, and scientific visualization. Such photorealistic images are generated by simulating the propagation of light in the real world, which is commonly called light transport simulation. In addition to the digital media, light transport simulation is also becoming a primary tool for analyzing the observed data (sensor measurements) of lights and training autonomous vehicles by generating photorealistic virtual street views. Light transport simulation computes the intensity of light arriving at the sensor (or the eye) given the input data of light sources and appearances and shapes of objects. Since light can bounce off objects any number of times, there are infinitely many possible paths of how light starting from a light source can arrive at the sensor. Light transport simulation commonly approximates such an infinite set of paths by randomly selecting several paths, and then estimates the intensity of light arriving at the sensor using those randomly selected paths. This approach is commonly called a Monte Carlo method. While widely used in the digital media industry, light transport simulation using this Monte Carlo approach is known to have several fundamental issues. The most commonly known issue is that it produces noisy images due to its numerical error. Since illumination is usually smooth, noise is an undesirable artifact that should be removed. It is also related to the fact that the Monte Carlo methods assume very little about the possible paths of light that indeed have many constraints. For example, both the energy of light and the intensity at the sensor are non-negative, but Monte Carlo does not explicitly use this fact. Monte Carlo approaches also ignore the recursive nature of light transport. For instance, illumination after three bounces should be similar to illumination after four bounces, but Monte Carlo approaches independently simulate different numbers of bounces. The objective of this research program is to challenge the common belief that Monte Carlo is the only viable approach for light transport simulation. Our goal is to address the fundamental issues of the application of Monte Carlo to light transport simulation by developing novel approaches that are impossible if we rely only on conventional Monte Carlo. My students and I work on different issues in parallel by exploring various approaches that do not fit within the paradigm of conventional Monte Carlo. This research program opens up a new paradigm of light transport simulation based on non-Monte Carlo approaches. The results contribute to the fundamental understandings of light transport simulation and lead to more efficient simulation of light transport. Such efficient light transport simulation is necessary to accommodate the ever-growing demands for visual details in the digital media industry.