Measuring the Perception of Visual Realism in Images

One of the main goals in realistic rendering is to generate images that are indistinguishable from photographs – but how do observers decide whether an image is photographic or computer-generated? If this perceptual process were understood, then rendering algorithms could be developed to directly target these cues. In this paper we introduce an experimental method for measuring the perception of visual realism in images, and present the results of a series of controlled human subject experiments. These experiments cover the following visual factors: shadow softness, surface smoothness, number of light sources, number of objects, and variety of object shapes. This technique can be used to either affirm or cast into doubt common assumptions about realistic rendering. The experiments can be performed using either photographs or computer-generated images. This work provides a first step towards objectively understanding why some images are perceived as photographs, while others as computer graphics.