Learning a Photorealism Score for 2D Images Extracted from a Railway Driving 3D Simulator

Abstract 3D modeling consists in describing objects thanks to a cloud of points in the three-dimensional space. It is ubiquitous nowadays and employed in many different fields such as video games, animation movies, architecture or medical applications. Subsequently, 2D images are retrieved from such 3D models through a process called 3D rendering. Recent hardware and software developments allow to approximate the real-world physics in an ever more precise manner. However, the question of the photorealism of such images remains. In this paper, we investigate in particular the photorealism of images extracted from 3D railway driving simulators. Two contributions are detailed in this article. First, we propose a CNN-based method designed to associate a photorealism score to any simulated image. It specifically quantizes the perceptual difference between computer-generated railroad images and natural ones. The second and main contribution of this work consists in the conception and implementation of a subjective experiment conducted so as to perceptually validate the aforementioned score. Human annotators are asked to specify which of two displayed images is more photorealistic, repeated over several pairs of images. After analysing the results, we conclude that a 96\% correlation exists between the proposed photorealism score and human visual perception.