The modeling of architectural structures is an important part of computer graphics, particularly with the growth in virtual environments for simulation and entertainment. Debevec, Taylor and Malik [1] recently developed techniques for reconstructing architectural models from a sparse set of photographs, and subsequently producing images of the model rendered from a novel viewpoint. The rendering algorithm derives the intensity of a point in the new image by combining the intensities for an appropriate set of points in the original photographs. This technique for generating pixel intensities has the disadvantage that the new rendering will inherit the illumination model from the original photographs. Not only does this limit the range of images that are possible, it also requires that the original photographs be taken in nearly identical lighting conditions.
We present a method for removing the effects of lighting on the original photographs, thus allowing them to be rendered under a different lighting model. In other words, we seek to recover the lambertian reflectance coefficient of every pixel in the original photograph. Rather than store this value, we produce an ambient image which depicts the scene as if on a cloudy day, and store the corresponding global ambient intensity. This method has the additional advantage that the resulting set of ambient images may be used to derive texture maps suitable for export to generic rendering programs. The lighting artifacts we remove are diffuse scattering effects and shadows caused by a known directional or point light source.