Graphics Turing Test

We define a Graphics Turing Test to measure graphics performance in a similar manner to the definition of the traditional Turing Test. To pass the test one needs to reach a computational scale, the Graphics Turing Scale, for which Computer Generated Imagery becomes comparatively indistinguishable from real images while also being interactive. We derive an estimate for this computational scale which, although large, is within reach of todays supercomputers. We consider advantages and disadvantages of various computer systems designed to pass the Graphics Turing Test. Finally we discuss commercial applications from the creation of such a system, in particular Interactive Cinema. Defining graphics computing scales can be difficult. Traditional methods such as counting triangles drawn per second can be misleading especially if a large percentage of the computing power is going into shading code or a physics simulation which can be involved in making the graphics image realistic. More than fifty years ago Alan Turing faced a similar challenge in evaluating the power of a computer. In short, he devised a test to see if one can build a computer whose intelligence was indistinguishable from human intelligence [1]. The Turing computational scale is then defined by the computational power required to pass this test. The beauty of this definition is that the computational scale is defined by the indistinguishability of comparison. In particular one does not need a definition of human intelligence to implement the test. Also one does not need to measure in terms of Flops per second or some other commonly used method of measuring computer power, after all integer performance, memory access or network connectivity may be important. One only needs to statistically compare how subjects can discriminate between a human and a computer. If the subjects can do no better than a random guess than the Turing test is passed and the Turing scale has been measured. Unfortunately the Turing Test is too hard for now, one cannot even use partial successes in AI to estimate the size of a computer required to pass the test [2]. As we shall see the corresponding situation in graphics is somewhat better.