Rating realism assessment for computer generated imagery

Computer-generated imagery (CGI) is becoming integral to a movie's story and appeal, and even dominates the film's success at box office. Currently the CGI realism is evaluated by post-production supervisors, and few objective realism assessments focus on this area. This paper investigates enhanced feature learning and classifier training for CGI assessment by deep learning. A training-set-selection method is proposed to select proper samples, which is crucial to deep learning training. Then the selected samples are converted into entropy images with enhanced features. We adopt a convolutional neural network for feature learning and classifier training to estimate the realism of CGI. Experimental results show that the developed matric has acceptable accuracy when compared to the grout truth. In addition, the rating result of the proposed assessment is very close to that of human visual perception.