Appearance models for computer vision

点击翻译按钮获取中文摘要

My research is concerned with fundamental questions about the visual appearance of the natural world.  How can we mathematically model visual cues such as shading, shadows, mirror reflections, texture, defocus, and motion parallax ?  How can we develop algorithms that are based on these models and that allow a vision system to make reliable inferences about 3D world, namely inferences about object layout and shape, surface material, and illumination.  The physics (i.e. optics) of image formation has been well studied and many models exist at  varying levels of simplicity.  The challenge from a computer vision point of view is to develop models that support reliable inferences from images.  The main approach I take in my research is to identify particular kinds of scenes and vision problems that have been neglected in the past.  For example, I am addressing the appearance of surface texture on highly slanted planes, such as at the boundary of a smooth object or at the horizon, and in particular I am examining how the appearance of texture depends on focus. Another problem is the relationship between shading and texture and how these cues can be distinguished in images.  A third problem is motion parallax which can occur in a variety of guises, included clutttered scenes such as forests, but also on curved mirror surfaces such as glass or metal.   In addition to addressing each of these problems in isolation, my research will also examine mathematical and computational relationships between these problems.