PERCEPTUAL CONSTRAINTS FOR IMAGE UNDERSTANDING IN HUMANS

当人们看着场景时，人们立即能够将边缘分类 该场景感知到该边缘是阴影边界，则该边缘是 曲面上的折痕等。如果我们认为视觉系统是 为了从图像数据解释场景属性，我们会说 视觉系统非常擅长于确定物理原因 图像中边缘投影到的轮廓。此外，视觉系统 使用图像中轮廓的配置提供的信息来 帮助解释场景属性，如曲面的形状。问题所在 包括这两个功能的等高线解释的出现 当一个人认为在一个图像中提供的信息 对于这两个函数，在局部上是不明确的。解决这个问题的方法 被许多计算机视觉研究人员采用的是指定一组 对等高线解释的约束，当它们结合在一起时，定义了 独特的解决方案。可以使用类似的方法来组织研究 人类对轮廓的感知处理。在这种方法中，人们考虑到 视觉系统隐含地在其 等高线的解释。 约束来自两个来源：图像数据本身和自然 环境的结构。本建议的目的是分析 这两种类型的约束可能对可视对象可用 等高线解释系统和从心理物理角度调查 视觉系统实际使用的约束的性质。这项研究 将重点放在已接收到的两种类型的轮廓上的约束 视觉研究人员相对较少的关注：反射率轮廓， 这是由于表面反射率和阴影的不连续性而产生的 轮廓，从场景中照明的不连续处投影。 也将考虑遮挡边缘，因为它们在 确定场景结构。研究的一个重要方面将是 使用3D渲染技术来生成自然图像 实验性刺激。这将允许操作独立的 3D场景域和2D图像域中的变量。一种能力 对场景的三维结构进行控制是进行场景研究的必要条件 视觉系统对边缘结构的自然约束是什么 因为这些是在场景域中指定的，而不是在图像域中指定的。这是 调查结果将使人们更深入地了解 图像中可用于等高线解释的信息以及 这些信息被人类的视觉系统所使用。更广泛地说， 这项研究涉及到多少环境知识的问题 结构和图像形成过程被结合到视觉中 系统处理。

When one looks at a scene, one is immediately able to classify the edges in the scene, perceiving that this edge is shadow boundary, that edge is a crease in a surface and so on. If we consider the visual system to be working to interpret scene attributes from image data, we would say that th visual system is very good at determining the physical causes of the contours to which edges project in an image. Moreover, the visual system uses information provided by the configuration of contours in an image to help interpret scene attributes such as the shapes of surfaces. The proble of contour interpretation, which includes both of these functions, appears quite difficult when one considers that the information provided in an imag for both functions is locally ambiguous. The approach to the problem adopted by many computer vision researchers is to specify a set of constraints on contour interpretation which, when taken together, define a unique solution. A similar approach can be used to organize research into human perceptual processing of contours. Within the approach, one consider the visual system as implicitly enforcing a set of constraints in its interpretation of contours. The constraints come from two sources; the image data itself and the natura structure of the environment. The present proposal aims to analyze the constraints of both types which are potentially available to the visual system for contour interpretation and to investigate psychophysically the nature of the constraints actually used by the visual system. The research will focus on constraints on two types of contour which have received relatively little attention from vision researchers: reflectance contours, which project from discontinuities in surface reflectance, and shadow contours, which project from discontinuities in illumination in a scene. Also considered will be occluding edges, due to their importance in determining scene structure. An important aspect of the research will be the use of 3D rendering techniques to generate naturalistic images for experimental stimuli. This will allow the manipulation of independent variables in both the 3D scene domain and the 2D image domain. The ability to control the 3D structure of scenes is necessary for the investigation of what natural constraints on edge structure are assumed by the visual system since these are specified in the scene domain, not in the image domain. Th results of the investigation will provide a deeper understanding of the information available in images for contour interpretation and the ways in which this information is used by the human visual system. More broadly, the research relates to the issue of how much knowledge of environmental structure and of the image formation process is incorporated into visual system processing.