While the promise of image-based rendering is great, the area has suffered from the lack of a formal problem definition. Most geometry-based rendering approaches can be accurately described as simulation problems. However, image-based approaches have lack any similar framework. Without a problem definition it is difficult to either judge, improve on, or compare results. I will present the "plenoptic function" of Adelson and Bergen as a concise problem statement for image-based rendering paradigms. This continuous 7-parameter function describes everything visible from any point in space. I will treat image-based rendering as a signal reconstruction problem where the plenoptic function is interpolated and extrapolated from a set of discrete samples.
In this talk I will present an image-based rendering approach based on a simple extension to the image-warping techniques described by Heckbert and Wolberg. This formulation has the advantage that it can be computed in an efficient incremental fashion well suited for hardware implementation. I will show how a two-dimensional image-flow field can be decomposed into a scalar function defined at each pixel. This scalar function is a generalized version of disparity used within the computer-vision community to compute depth-from-stereo. This disparity value can also be used to perturb a traditional perspective warping function of a plane. This warping amounts to the transformation of the original image to new viewpoint.
Finally, I will discuss open problems, applications, and system implications of image-based computer graphics.
Host: Prof. Seth Teller
|
Modified: Jun 25, 1997|
Current events|
Your comments
and inquiries are welcome.