A survey on generative 3D digital humans based on neural networks: representation, rendering, and learning

Abstract

With the rapid development of artificial intelligence, interdisciplinary research addressing computer vision and graphics has engendered a revolution in digital human generation. The dream of human beings entering the “metaverse” and other digital spaces is gradually becoming a reality. Traditional graphics-based modeling has cumbersome requirements and a long cycle, which is no longer optimal for large-scale digital human generation. Therefore, a popular research topic has been to learn generative models for generating high-fidelity digital humans. In this survey, we review digital human technology from the perspective of generative models and analyze the research status and status quo of 3D digital human technology. We first summarize three major components in the pipeline, namely, model representation, rendering, and learning. Then, we summarize the modeling methods based on explicit and implicit representation. We also analyze traditional rendering and neural rendering approaches and the corresponding learning methods. Finally, we discuss the typical applications of 3D digital humans and summarize current challenges and future research directions.