Visual Perception as Statistical Inference

视觉感知作为统计推断

• 批准号: 9631682
• 负责人: Daniel Kersten
• 金额: $ 20.42万
• 依托单位: University of Minnesota-Twin Cities
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 1996
• 资助国家: 美国
• 起止时间: 1996-10-01 至 2000-10-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=9631682&HistoricalAwards=false
• 关键词: Visual; Perception; Statistical; Inference

9631682 KERSTEN One of the great mysteries of science is how the human visual system determines what and where objects are just by looking. When viewing a scene, the visual system solves, in an instant, the problems of recognizing objects, deciding what they are made of, and where they are relative to each other and the viewer. The seeming effortlessness of this achievement is all the more remarkable because the image input to the eye is locally ambiguous, i.e., in any patch of an image, there is rarely an unequivocal cue to how far away it is, what it is made of, or how it is illuminated. Somehow, the brain makes its best bet as to what is out there from this ambiguous image, and these decisions are rarely wrong. In addition, the image of an object is highly variable depending on illumination and viewpoint. For example, illuminating an object from the left or right has relatively little effect on our perception of the object, despite the fact that the object's two images are quite different. This research stems from the hypothesis that vision solves the problem of ambiguity by exploiting the constraints on how images are formed and on the a priori statistical regularities inherent in object and scene structure. From this point of view, vision is a process of statistical inference that makes good bets based on available information and solves the problem of variability, in part, by constraints that are determined by the task itself, such as determining where, rather than what, an object is. This research will investigate human perception of the spatial layout of objects and object properties. Vision determines spatial layout from well over a dozen sources of depth information, including cast shadows. It is well-known to artists that the closer the image of an object's shadow is to the object itself, the closer the object is to the surface receiving the shadow. How visual perception makes this inference is not clear, because local image informatio n does not uniquely specify whether a local brightness change is a shadow or not. This research will identify constraints in the image and scene that resolve the ambiguities in the perception of depth from shadows. As for the perception of spatial layout, determining object properties and identity is also confounded by local ambiguity as well as by variability in illumination. Experiments comparing the performance of human and statistically optimal observers for inferences about object shape, material and identity will tease apart the relative roles of prior knowledge and image cues in removing uncertainty about the properties of objects. Visual perception is a major way in which we acquire knowledge of the world. Computer scientists have yet to produce a machine that can recognize objects in natural images. Understanding how human vision resolves ambiguity about object location and identity promises to guide the development of artificial vision systems and help to unravel the mysteries of the brain itself. ***

9631682 KERSTEN 科学的最大谜团之一是人类视觉系统如何通过观察来确定物体是什么以及在哪里。 当观看场景时，视觉系统会立即解决识别物体、确定物体由什么构成以及它们彼此之间以及相对于观看者的相对位置的问题。 这一成就看似毫不费力，却更加引人注目，因为输入到眼睛的图像局部是模糊的，即在图像的任何一块中，很少有明确的线索表明它有多远，它是由什么构成的，或者它是如何被照亮的。 不知何故，大脑对这个模糊图像中的内容做出了最好的选择，而且这些决定很少是错误的。 此外，物体的图像根据照明和视点的不同而变化很大。 例如，从左侧或右侧照亮一个物体对我们对该物体的感知影响相对较小，尽管该物体的两个图像有很大不同。 这项研究源于这样的假设：视觉通过利用图像形成方式以及物体和场景结构固有的先验统计规律的约束来解决模糊问题。 从这个角度来看，视觉是一个统计推断的过程，它根据可用信息做出正确的选择，并部分地通过任务本身确定的约束来解决可变性问题，例如确定在哪里，而不是什么，一个物体是。 这项研究将调查人类对物体空间布局和物体属性的感知。 视觉通过十多个深度信息源（包括投射阴影）来确定空间布局。 艺术家们都知道，物体的阴影图像越接近物体本身，物体就越接近接受阴影的表面。 视觉感知如何做出这一推论尚不清楚，因为局部图像信息并不能唯一地指定局部亮度变化是否是阴影。 这项研究将确定图像和场景中的约束，以解决阴影深度感知的模糊性。 至于空间布局的感知，确定对象属性和身份也会受到局部模糊性以及照明变化的影响。 比较人类和统计上最佳观察者在推断物体形状、材料和身份方面的表现的实验将梳理先验知识和图像线索在消除物体属性不确定性方面的相对作用。 视觉感知是我们获取世界知识的主要方式。 计算机科学家尚未生产出能够识别自然图像中的物体的机器。 了解人类视觉如何解决物体位置和身份的模糊性有望指导人工视觉系统的开发，并有助于揭开大脑本身的奥秘。 ***