An experimental evaluation of computer graphics imagery

Reviewer: Paolo E. Sabella

Two experiments are presented in which comparisons are made between computer generated images of computer models and the physical environments that the models represent. The authors' objective is to test the accuracy, both measured and perceptual, of their synthesis of realistic scenes. They are dealing with diffuse environments, a class of scenes for which standard image synthesis techniques are particularly deficient. The first experiment compares, for three different environments, radiometric measurements made of the radiant energy flux densities with the computed predictions based on the radiosity method. The model of light propagation takes into account the measured spectral reflectance of materials in the scene and the spectral and directional characteristics of the light source. The results show agreement to within ±7 percent. The second experiment compares the perception of the real and synthesized scenes by subjects in a controlled setting. The setup consisted of placing a view camera pointing at the real scene next to another similar camera pointing at a color monitor displaying the synthesized image. It is worth noting that the computer image is compared with light emitted from a real scene, not with a televised image of the scene. The results are that 45 percent of the subjects selected the wrong answer, about the same as they would have by guessing. While the radiosity method copes with diffuse and interbody illumination allowing a computation of “bleeding,” the spectral accuracy of the method was not demonstrated by these experiments due to the type of radiometer used. (The authors do concede in the Conclusion that this is an area for further research.) This paper does demonstrate that realism of image synthesis can be experimentally validated. It is a start of what may evolve into more routine techniques for obtaining benchmarks of realism among competing image synthesis methods.