Is Seeing Still Believing: Factors That Allow Humans and Machines to Discriminate Between Real and Generated Images

The growing photorealistic capabilities of 3D animation and computer-generated imagery (CGI) poses potential issues for broadcasters, regulators, governments, and viewers. What happens if a news story is broadcast, or available on a website, that is CGI or a combination of CGI and camera-acquired images? There are already instances of video where there is polarized debate as to the degree of CGI present in the published content, a prime example being the Islamic State of Iraq and Syria (ISIS) video of the burning of a person. An analysis of the artefacts seen in the video that gave rise to the truthfulness debate is presented. The factors that humans use to discriminate or to evaluate the realness of photographs include shadow softness, surface smoothness, scene complexity and composition, and the number of light sources. Initial studies showed a human ability to discriminate; however, recent studies have shown that the ability to discriminate is getting more difficult, although with training the ability improves. Computer Vision research has evaluated image features and descriptors that discriminate between photographs and CGI. No such studies with quantitative data or experimental methodologies seem to exist for evaluating the CGI moving image by humans or computers. Possible temporal assessment factors for human discrimination (motion parameters), for computer vision (optical flow), and for artificial intelligence (semantic scene analysis) are presented.