CPA -G&V: Collaborative Research: Visual Equivalence: a New Foundation for Perceptually-Based Rendering of Complex Scenes

摘要：卡维塔·巴拉(0811680)真实感成像的目标是生成逼真地代表真实世界的图像。评估图像保真度是真实感成像的一个重要方面，在计算机图形学和数字摄影中都有广泛的意义。最近，基于人类视觉计算模型的感知度量被用于表征图像的保真度，但这些度量通常表征像素级的图像差异。当我们看图像时，我们看到的不是像素，而是具有不同形状、大小、材质和运动的物体。因此，需要表征对象的外观属性的图像保真度度量。该项目引入了一种新的图像保真度标准，称为视觉等效性。如果图像传达了关于对象外观(形状、材质、照明)的相同信息，即使它们逐个像素不同，它们在视觉上也是等效的。这一新标准从根本上扩展了如何评估图像保真度的范围，提供了一种基于视觉编码的更高级别属性的度量。本项目试图通过以下方式来理解和利用视觉等值现象：a)进行一系列心理物理实验，通过探索复杂场景中对象的几何、材质和照明属性之间的相互作用来为视觉等值奠定基础；b)开发适用于广泛场景的视觉等值预测器；以及c)展示视觉等值概念在一系列应用中的实用价值，包括高保真环境地图压缩和复杂场景的可伸缩渲染。这项研究的成果为解决计算机图形学中的一个重大挑战提供了有价值的新工具：复杂场景的高性能、高保真图像合成。此外，这种新的图像保真度方法应该在图形以外的领域产生重大影响，包括2D和3D数字图像的获取、编码、压缩、传输、存储和显示。

AbstractPI: Kavita Bala (0811680)The goal of realistic imaging is to produce images that faithfully represent the appearance of the real world. Assessing image fidelity is an important aspect of realistic imaging that has broad implications in both computer graphics and digital photography. Recently perceptual metrics based on computational models of human vision have been used to characterize image fidelity; but these metrics typically characterize pixel-wise image differences. When we look at images we don?t see pixels, we see objects with distinct shapes, sizes, materials, and motions. Therefore, there is a need for image fidelity metrics that characterize the appearance properties of objects. This project introduces a new standard for image fidelity called visual equivalence. Images are visually equivalent if they convey the same information about object appearance (shape, material, lighting), even if they are different pixel-by-pixel. This new standard fundamentally expands the scope of how to assess image fidelity, providing a metric that is based on higher-level properties of visual coding.This project seeks to understand and leverage the phenomenon of visual equivalence by: a) conducting a series of psychophysical experiments that develop a foundation for visual equivalence by exploring interactions between the geometry, material and illumination properties of objects in complex scenes; b) developing visual equivalence predictors that apply to a wide range of scenes; and c) demonstrating the practical utility of the concept of visual equivalence in a range of applications including high-fidelity environment map compression and scalable rendering of complex scenes. The products of this research provide valuable new tools for addressing one of the grand challenges in computer graphics: the high performance, high fidelity image synthesis of complex scenes. Further, this new approach to image fidelity should have a significant impact beyond graphics in areas including 2d and 3d digital image acquisition, coding, compression, transmission, storage, and display.