ABSTRACT

An image depends on a number of scene properties, including shape, surface reflectance, scene illumination and the observer's viewpoint. These properties interact in a complex manner, often creating specular highlights and other intricate visual effects. Understanding and modeling these effects is important for computer vision systems to succeed in uncontrolled environments. In the past, it has been common in computer vision to use rather simple models of appearance, involving such things as perfectly matte surfaces and isolated point light sources. While these models lead to tractable inference algorithms (shape from shading is a good example), they are rather crude and often lead to systems that perform poorly in “real-world” settings. This talk is meant to convey an alternative approach. I argue that although our visual world is diverse and complex, there are properties of materials and illumination that are common to many scenes. By developing computational tools for exploiting these general properties, we can utilize image data more efficiently and improve the performance of many vision systems. Along these lines, I will present new computational approaches to a few visual tasks, including color constancy and the inference of 3D shape from photometric cues. Biography: Todd Zickler received his Ph.D. in electrical engineering from Yale University in 2004 and is currently an assistant professor in the School of Engineering and Applied Sciences at Harvard University. His research spans computer vision, computer graphics and image processing. He is currently focused on developing representations of visual appearance and ways to exploit them for visual inference. In 2006, he was the recipient of an NSF career award titled, “Foundations for Ubiquitous Image-based Appearance Capture.” Reception at 4:30 in OLS 224 (Student Lounge) Other Recent and Upcoming Colloquia