This paper is at the boundary of computer graphics and computer vision. As such, it addresses a problem where the goal is to generate a scene that is both real (the computer vision problem) and attractive-looking (the computer graphics problem).
The goal of the work is to recover the complete reflectance information for the surfaces in a scene, using a single image, in order to use graphics techniques to manipulate the final result. This will make it possible to add objects and change the viewpoint or lighting.
The paper gives a good review of the various approaches that have been used to solve parts of this problem. The authors detail the techniques they have used to generate a more complete solution and provide one example. Their solution iteratively compares its prediction with the actual image and applies a more complex model to the regions where the differences are too large.
This paper presents an important contribution in both computer vision and computer graphics. Unfortunately the space limitations constrained the size of most of the figures, which makes some of them difficult to read.