Stereoscopic Dataset from A Video Game: Detecting Converged Axes and Perspective Distortions in S3D Videos

Abstract

This paper presents a method for generating stereoscopic or multi-angle video frames using a computer game (Grand Theft Auto V). We developed a mod that captures synthetic frames allows us to create geometric distortions like those that occur in a real video. These distortions are the main cause of viewer discomfort when watching 3D movies. Datasets generated in this way can aid in solving problems related to machine-learning-based assessment of stereoscopic- or multi-angle-video quality. We trained a convolutional neural network to evaluate perspective distortions and converged camera axes in stereoscopic video, then tested it on real 3D movies. The neural network discovered multiple examples of these distortions.