The removal of lighting artifacts requires knowledge of the illuminants in the scene. In theory this information could be derived from information within the scene, such as the orientation and location of shadow boundaries, or specular reflection. However, such features are difficult to detect within an image without human help. Hence we assume that the location of the light source is known, or sufficient information to derive it is available. To this end, a user has the option of specifying a date, time and location for the photograph from which sun position can be calculated [2].
There is insufficient information at a single pixel location to determine the BRDF at that point. We hence make the assumption that the surface is lambertian. However, even with this assumption, there is insufficient information to determine the reflectance coefficient of a single point. One solution is to identify points that should have the same reflectance, but with a textured surface this is a difficult problem. Furthermore, the pixels chosen must be lit under different conditions (either different incident light directions or different shadow states). This precludes the choice of pixels from a single fully lit or shadowed face. To overcome these problems, we choose to assume that all the faces of a given block in the model have the same average reflectance. In general we will always be able to see two faces of at least one block under different conditions, allowing us to extract the ambient intensity for the photograph and remove lighting artifacts. Note that a user has some control over this assumption through the block structure of the model.
Finally, we assume that the ambient conditions are uniform across the original photographs. This assumption is not essential, but helps to smooth out other approximations and numerical errors.