The process described above was tested on a set of photographs of the Campanile tower at UC Berkeley. For comparison purposes, a single photograph (fig 1) was processed to produce an ambient image (fig 2). This ambient image was then used as the basis for a re-rendering under the same lighting conditions as the original model. The result is shown in figure 3.
Figure 1: The original Campanile photograph.
Figure 3: A re-rendering of the original photograph.
The ambient image appears to be largely accurate, in that it resembles the building on a cloudy day. The most significant deviation is on the right side of the main tower at the top, where the faces appear unusually bright. This is most likely due to a combination of differences in the average block reflectance for the shadowed and visible face. There is also some interaction with the shadow computations, most evident on the lower side of the narrow ledge halfway up the image.
The re-rendering closely matches the original photograph in most places. The large frustum at the top is particularly well rendered. The success in this region is due to the fact that it most closely satisfies the assumptions of constant reflectance across a block. However, the tower appears to be too bright on the right side near the top. This is a combination of the aforementioned errors in the ambient image and quantization of pixel brightnesses. Note also that the windows into the bells have a different appearance to the left and right in the original image. This difference is largely responsible for the violation of the constant reflectance function. It could be alleviated by more accurately modeling the true geometry of the windows.
As a final example of the Campanile images, figure 4 depicts the same model under different lighting conditions. The intensity of the light source has been modified, and the direction has been changed such that the light now comes from the lower left side of the image. Note that the left side is now brighter than the right side. Note also the shadows on the small cap at the top of the right side of the tower.
Figure 4: A novel rendering under new lighting conditions.
The algorithm was also tested on images that were taken on cloudy days. In this case, the ambient image is the same as the original photograph. Figures 5 and 6 depict a school building model under 2 different lighting conditions, neither of which matches that of the original photographs. The results are somewhat better in this case, because there are no inaccuracies introduced by the processing of images into ambient images. Figure 5 depicts the building with the sunlight coming in from behind the building. Note the large shadow the building casts on the ground-plane, and the various areas of self shadow. Figure 5 represents the same viewpoint, but with the sun to the right of the building.
Figure 5: The University School building rendered with the sun from behind.
Figure 6: The University School building with sunlight from the right.
