苏黎世联邦理工学院：从视觉和语言学习数字人类（英文版）（248页）.pdf

1、Learning Digital Humans fromVision and LanguageDiss.ETH No.30694Yao FengDiss.ETH No.30694Diss.ETH No.30694Learning Digital Humans fromVision and LanguageA thesis submitted to attain the degree ofDoctor of Sciences(Dr.sc.ETH Zurich)presented byYao FengMaster of Engineering in Electronics and Communic

2、ation Engineering,Shanghai Jiao Tong UniversityBorn on 21.01.1995accepted on the recommendation ofProf.Dr.Marc PollefeysProf.Dr.Michael J.BlackProf.Dr.Fernando De la Torre Frade2024iiAbstractThe study of realistic digital humans has gained significant attention withinthe research communities of comp

3、uter vision,computer graphics,and ma-chine learning.This growing interest is driven by the importance of under-standing human selves and the pivotal role digital humans play in diverseapplications,including virtual presence in AR/VR,digital fashion,enter-tainment,robotics,and healthcare.However,two

4、major challenges hinder the widespread use of digital hu-mans across disciplines:the difficulty in capturing,as current methods relyon complex systems that are time-consuming,labor-intensive,and costly;and the lack of understanding,where even after creating digital humans,gaps in understanding their

5、 3D representations and integrating them withbroader world knowledge limit their effective utilization.Overcomingthese challenges is crucial to unlocking the full potential of digital humansin interdisciplinary research and practical applications.To address these challenges,this thesis combines insi

6、ghts from computervision,computer graphics,and machine learning to develop scalablemethods for capturing and modeling digital humans.These methods in-clude capturing faces,bodies,hands,hair,and clothing using accessibledata such as images,videos,and text descriptions.More importantly,wego beyond cap