Synthesizing Environment Maps from a Single Image

Environment mapping is a popular technique for creating consistent lighting when compositing a virtual object into a real scene. However, capturing an environment map usually requires physical access to the scene to obtain illumination measurements. But what if all one has available is a single photograph of the scene? In this paper, we study techniques for synthesizing plausible environment maps from a single image. By analogy with texture synthesis, the goal is to use the small amount of available data to generate an environment map that would be likely to have come from that scene. In particular, we are interested in understanding the role of geometric information in constructing visually realistic environment maps. To this end, we implement several environment synthesis strategies that employ varying amounts of 3D scene geometry information. We measure the quality of the synthesized results by using human subjects to evaluate the appearance of objects illuminated with different environment maps, in still images as well as in video.