Human visual perception of shape, lightness, and lighting

点击翻译按钮获取中文摘要

What could be easier than seeing? We see without trying, and usually without thinking about it. Seeing seems easy because our visual cortex is a powerful computing device with a lifetime of experience at seeing things. One reason why vision is a challenging computational task, though, is that all visual images are highly ambiguous, and any given image could conceivably be seen as depicting any of a wide range of shapes, colours, and lighting conditions. Our visual system needs to overcome this ambiguity somehow, in order to perceive scenes and objects correctly. One way our visual system overcomes ambiguity is by relying on assumptions about what shapes, colours, and lights are most likely to occur in the scenes we see. For example, previous experiments have shown that starting from infancy, our visual system slowly learns that light usually shines from overhead, and our visual system uses this knowledge to perceive objects and scenes correctly. The main question addressed by the present research program is, what other facts about real-world shapes, colours, and lighting does the human visual system rely on in order to perceive the world correctly? We will investigate this question in two ways. First, using novel experimental tools called classification images, we will investigate what parts of visual images affect peoples' judgements about the brightness, orientation, or lighting conditions at a target location in an image. We will do this by introducing small random fluctuations into visual images, and over many thousands of presentations, measuring which random fluctuations affect peoples' judgements and which do not. We will use the results of these experiments to test computational theories of how people perceive shape, colour, and lighting. Second, we will build a new computational model that makes statistical assumptions about the shapes, colours, and lighting conditions in the scenes we see, and that uses these assumptions to interpret visual images. We will investigate what assumptions such a computational model must make, in order to interpret visual stimuli in ways that are as similar as possible to what human observers see. This research will help us understand normal human vision, both in day-to-day life and in more difficult tasks like operating in immersive environments or judging medical images. It will also provide information that will be useful for developing computer vision systems that see scenes and objects in the same way that people do.