Bayesian shape from shading

描述（由申请人提供）：这项研究的长期目标是了解使人类观察者能够感知物体形状的视觉机制。这些机制是正常视觉功能的核心，理解它们将有助于阐明正常视觉的原理和局限性，以及导致视觉缺陷，特别是视觉失认症的缺陷。这些实验的具体目的是研究人类观察者从视网膜图像中的阴影信息中感知物体形状的能力。这种阴影信息是非常模糊的，因为任何给定的2D图像都可以由无限数量的光源和表面的3D排列中的任何一个生成。然而，人类观察者很快就能对这些图像做出大致准确的解释，这意味着人类视觉系统必须包含准确的假设（至少是隐含的），即哪些3D场景更有可能或更少出现在现实世界中。这里测试的假设是，视觉系统使用自然光照的统计模型和局部表面形状和反射率的统计模型，以便最可能地对2D图像进行3D解释。这个研究项目有四个组成部分。(1)利用多向光度计，在多个随机选择的地点测量来自各个方向的光照入射，并利用这些测量值构建自然光照的统计模型。(2)利用三维数字化扫描仪，将许多自然物体的形状和反射率模式数字化，并利用从这些数据中提取的小表面斑块构建局部表面形状和反射率的统计模型。(3)我们将开发一个贝叶斯模型观测器，该观测器使用测量的照明和表面统计数据来推断2D图像中描绘的物体最可能的3D形状。(4)我们将比较人类观察者和贝叶斯模型观察者在一些形状感知任务上的表现。如果人类和贝叶斯观察者发现相同的任务容易或困难，那么我们将得出结论，人类的表现受到与贝叶斯观察者表现相同的因素的限制，即各种任务中的照明和表面统计与自然世界中测量的统计相匹配的程度。否则，人类观察者必须使用测量到的自然统计数据以外的约束来对2D图像进行精确的3D解释。

DESCRIPTION (provided by applicant): The long-term goal of this research is to understand the visual mechanisms that enable human observers to perceive the shapes of objects. These mechanisms are central to normal visual functioning, and understanding them would help to illuminate principles and limitations of normal vision, as well as the deficiencies that cause visual deficits, especially visual agnosias. The specific aim of these experiments is to investigate human observers' abilities to perceive the shapes of objects from the shading information in retinal images. This shading information is extremely ambiguous, because any given 2D image could have been generated by any of an infinite number of 3D arrangements of lighting sources and surfaces. Nevertheless, human observers rapidly arrive at a generally accurate interpretation of such images, which implies that the human visual system must incorporate accurate assumptions (at least implicitly) about which 3D scenes are more or less likely to occur in the real world. The hypothesis being tested here is that the visual system uses a statistical model of natural illumination and a statistical model of local surface shape and reflectance, in order to arrive at the most likely 3D interpretation of a 2D image. This research project has four components. (1) Using a multidirectional photometer, we will measure the illumination incident from all directions at many randomly selected locations, and use these measurements to construct a statistical model of natural illumination. (2) Using 3D digitizing scanners, we will digitize the shape and reflectance patterns of many natural objects, and use small surface patches extracted from this data to construct a statistical model of local surface shape and reflectance. (3) We will develop a Bayesian model observer that uses the measured illumination and surface statistics to infer the most likely 3D shape of objects depicted in 2D images. (4) We will compare the performance of human observers and the Bayesian model observer on a number of shape perception tasks. If the human and Bayesian observers find the same tasks easy or difficult, then we will conclude that human performance is limited by the same factor as the Bayesian observers performance, namely the extent to which the illumination and surface statistics in the various tasks match the statistics measured in the natural world. Otherwise, human observers must use constraints other than the measured natural statistics to arrive at accurate 3D interpretations of 2D images.