Assessing the Qualities of Synthetic Visual Data Production

Abstract

A literature review was conducted using journal articles and conference proceedings to examine emerging research practices, and applications of synthetic visual data over the past 5 years. The current research examined articles related to research trends in artificial intelligence training intended to improve computer vision and object detection. Search strings were developed and used to retrieve research articles from the ACM and IEEE databases. The resulting articles were examined for trends, general practices, disciplines where the greatest efforts have been made, advances, and relevant production processes. The research reveals that visual synthetic data encompasses filtering, augmentation, and object domain randomization techniques. Further, all of the research that included an evaluation of synthetic visual data suggest that there are noteworthy performance improvements in accuracy. Additionally, producing realistic synthetic data reduces the current limitations related to labeling, image quality, paucity of relevant data, and privacy issues.