大数据 ›› 2023, Vol. 9 ›› Issue (3): 114-139.doi: 10.11959/j.issn.2096-0271.2022081

• 研究 • 上一篇    下一篇

虚拟人形象合成技术综述

邓钇敏1,2, 张旭龙1, 司世景1,3, 王健宗1, 肖京1   

  1. 1 平安科技(深圳)有限公司,广东 深圳 518063
    2 中国科学技术大学,安徽 合肥 230026
    3 上海外国语大学国际金融贸易学院,上海 200083
  • 出版日期:2023-05-15 发布日期:2023-05-01
  • 作者简介:邓钇敏(1999- ),女,中国科学技术大学硕士生,中国计算机学会会员,主要研究方向为深度学习、计算机视觉、元宇宙等。
    张旭龙(1988- ),男,博士,平安科技(深圳)有限公司高级算法研究员,主要研究方向为语音合成、语音转换、音乐信息检索、机器学习和深度学习方法在人工智能领域应用。
    司世景(1988- ),男,博士,平安科技(深圳)有限公司资深算法研究员,深圳市海外高层次人才。美国杜克大学人工智能博士后,中国计算机学会会员,主要研究方向为机器学习和及其在人工智能领域应用。
    王健宗(1983- ),男,博士,平安科技(深圳)有限公司副总工程师,资深人工智能总监,联邦学习技术部总经理。美国佛罗里达大学人工智能博士后,中国计算机学会高级会员,中国计算机学会大数据专家委员会委员,曾任美国莱斯大学电子与计算机工程系研究员,主要研究方向为联邦学习和人工智能等。
    肖京(1972- ),男,博士,中国平安集团首席科学家,2019年吴文俊人工智能杰出贡献奖获得者,中国计算机学会深圳分部副主席,主要研究方向为计算机图形学学科、自动驾驶、3D显示、医疗诊断、联邦学习等。
  • 基金资助:
    广东省重点领域研发计划“新一代人工智能”重大专项”(2021B0101400003)

Human avatars synthesis technologies: a survey

Yimin DENG1,2, Xulong ZHANG1, Shijing SI1,3, Jianzong WANG1, Jing XIAO1   

  1. 1 Ping An Technology (Shenzhen) Co., Ltd., Shenzhen 518063, China
    2 University of Science and Technology of China, Hefei 230026, China
    3 School of Economics and Finance, Shanghai International Studies University, Shanghai 200083, China
  • Online:2023-05-15 Published:2023-05-01
  • Supported by:
    The Key Research and Development Program of Guangdong Province(2021B0101400003)

摘要:

随着元宇宙兴起,针对虚拟人形象化高效建模的需求日益迫切。从人类图像数据集中构建人类模型一直是计算机视觉的热门话题,其中3D虚拟人合成可以视作三维重建的子模块,重点在于对复杂的人体结构和表面细节的还原。对近年来虚拟人形象构建相关文献进行了全面调研,研究范围覆盖了全身形象、头部形象以及衣物建模等领域。分析归纳构建工作的基本原理,从各自技术路线层面出发将虚拟人合成方法分为基于网格、基于图像、基于体素、基于隐式表示、混合表示5类。首先介绍各类方法的基本原理,然后结合现有工作讨论具体技术,并指出各类方法的优缺点。此外还介绍了部分常见的模型质量评估的数据集和评价指标,简要介绍了虚拟人的常见应用。最后对虚拟人合成技术未来发展方向进行了展望,以合成高质量、高保真度、低延迟的虚拟人形象。

关键词: 元宇宙, 虚拟人, 三维人体重建, 计算机视觉, 深度学习, 人脸合成

Abstract:

Nowadays, the demand for efficient human avatars modeling is becoming increasingly urgent since metaverse has attracted more and more attention.Creating human avatars from human image datasets has always been a popular topic in the field of computer vision.3D human avatars synthesis can be regarded as a sub-module of 3D reconstruction focusing on reproducing the complex articulated body and surface details of human.A comprehensive survey of the literature related to the human reconstruction in recent years was conducted, including the work of full-body avatars, talking-head and clothing modeling.By analyzing and summarizing existing work, human avatars synthesis technologies were divided into five categories: mesh-based methods, image-based methods, voxel-based methods, implicit methods and hybrid methods due to the features of their pipelines.Firstly, the basic principles of them were introduced respectively.Secondly, the realization based on related work was discussed and then the advantages and disadvantages of methods respectively were pointed out.Thirdly, the datasets and metrics for model quality evaluation were introduced.Besides, an overview of various applications was given.Finally, the future directions of human avatars synthesis technology were prospected to synthesize high-quality, high-fidelity and low-latency human avatars.

Key words: metaverse, human avatars, three-dimensional human reconstruction, computer vision, deep learning, face synthesis

中图分类号: 

No Suggested Reading articles found!