基于神经网络的生成式三维数字人研究综述:表示、渲染与学习

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

随着人工智能技术的高速发展,计算机视觉与图形学等相关学科的交叉融合掀起了一场数字人生成技术的新革命,人类进入“元宇宙”等数字空间的梦想正逐渐变为现实.面对大规模三维数字人的生产需求,传统图形学建模方法建模过程繁琐,周期冗长,阻碍了虚拟数字人的普及和应用,而利用生成式人工智能技术产生高拟真、规模化的虚拟数字人正逐渐成为研究热点.为了深入了解三维数字人技术的研究现状与挑战,本文从生成式模型的视角对数字人技术进行了系统性梳理,并总结了其中的3个关键步骤:表示、渲染与学习.随后,对显式及隐式的表示方法进行总结,对传统渲染与神经网络渲染的成像方式进行归纳,并概括了相应的模型学习方法.最后,本文对三维数字人的典型应用进行分析,并对当前挑战与未来发展方向进行总结和展望.

With the rapid development of arti cial 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-delity 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 rst 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.