Collaborative Research: HCC: Medium: Neural Materials for Realistic Computer Graphics

真实感一直是计算机图形学的重要目标，不仅因为它使图像更具说服力，使虚拟环境更具沉浸感，还因为当使用渲染来做出现实世界的决策时，准确性很重要。真实感图像是通过使用模型来模拟从场景中的表面反射光线的物理过程而生成的，而最先进的反射模型虽然对于平滑、均匀的材质是准确的，但对于具有详细表面结构的材质来说却非常不准确。这个项目探索了一种全新的建模表面反射的方法，通过从表面的许多详细测量中学习特定类别材料的特征反射模式。这些新模型将产生广泛的影响，彻底改变在电影制作、工业设计、营销、广告、建筑、虚拟现实和房屋改建等应用中为计算机图形建模的材料的方式。这项研究围绕构建更适合描述精细细节的模型展开。当前的“基于物理的材质”本质上是参数反射模型，其参数由纹理贴图进行调制。这些可以匹配聚合的大规模行为，但不能匹配近距离真实反射的复杂性。这个项目探索了一种基于神经函数逼近器的表示材料的全新方法。这项工作将不仅仅是简单地将反射系数附加到不透明的物体上，而是通过开发适应模糊、半透明和未建模几何细节的薄神经反射率场。新的神经材料将是灵活和通用的，适用于从虚拟环境到视觉效果所需的所有材料，支持的应用范围从仅有的几张参考照片的自动外观开发，到结合几何和材料细节的图像反向对象建模。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Realism has always been an important goal of computer graphics, not only because it makes images more convincing and virtual environments more immersive, but because accuracy is important when renderings are used to make real-world decisions. Realistic images are made by using models to simulate the physical process of light reflection from surfaces in the scene, and state-of-the-art reflection models, while accurate for smooth, homogeneous materials, are quite inaccurate for materials with detailed surface structure. This project explores an entirely new way of modeling surface reflection, by learning the characteristic reflectance patterns of particular classes of materials from many detailed measurements of surfaces. These new models will have broad impact by completely transforming the way materials are modeled for computer graphics in applications like moviemaking, industrial design, marketing, advertising, architecture, virtual reality and home remodeling.This research centers around building models that are more appropriate for describing fine-scale detail. Current “physically based materials” are essentially parametric reflectance models with parameters modulated by texture maps. These can match aggregate large-scale behavior, but not the complexities of real reflectance at close scale. This project explores a fundamentally new way to represent materials, based on neural function approximators. The work will go beyond simply attaching reflectance to opaque objects, by developing thin neural reflectance fields that accommodate fuzziness, translucency and unmodeled geometric detail. The new neural materials will be flexible and general for the full range of materials needed from virtual environments to visual effects, supporting applications ranging from automatic look development from only a few reference photographs, to inverse object modeling from images, combining geometric and material detail.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.