【GAMES Webinar 2024-337期】
视觉专题
弱约束视觉重建：从数字人到物理仿真
· 1 ·
报告题目
PuzzleAvatar: Assembling 3D Avatars
from Personal Albums
报告嘉宾
Yuliang Xiu
MPI-IS
报告时间
2024年8月29号 晚上8:00-8:45（北京时间）
报告方式
GAMES直播间: https://live.bilibili.com/h5/24617282
报告摘要
PuzzleAvatar represents a paradigm shift in image-based 3D clothed human reconstruction. It uses casual, TRULY unconstrained photos—like those from your personal album—to create 3D avatars without requiring ANY re-projection photometric losses or geometric regularizers. We refer to this task as "Album2Human," which is extremely challenging due to the diverse poses, challenging viewpoints, cropped views, and random occlusions found in the daily captured casual photos. Yet PuzzleAvatar can swallow this unstructured data and deliver a structured output—the A-posed avatar—with a faithful appearance (shape+color), all while bypassing the challenging estimation of body pose for canonicalization, and camera pose for synchronization. In short, PuzzleAvatar can easily scale up with ALL accessible daily photos to create your digital twins.
嘉宾简介
Yuliang Xiu is currently a final-Year Ph.D. in Perceiving Systems, Max Planck Institute for Intelligent Systems (MPI-IS), working with Michael J. Black. His research interest mainly lies in the intersection of computer vision and computer graphics, especially in 3D human digitalization. He was briefly enrolled (2019-2020) as a Ph.D. student at the University of Southern California before he relocated to Tübingen, advised by Hao Li. He will join Westlake University as a tenure-track Assistant Professor in Spring 2025, and lead Endless AI Lab (endless.do) as PI.
个人主页
https://xiuyuliang.cn
· 2 ·
报告题目
Next-Gen Camera System: Instilling
Intuitive Imagination in Smartphones
报告嘉宾
官善琰
vivo影像规划中心
报告时间
2024年8月29号 晚上8:45-9:30（北京时间）
报告方式
GAMES直播间: https://live.bilibili.com/h5/24617282
报告摘要
AI尤其是Generative AI极大地拓宽了相机的摄影能力和可玩性，但仍未能赋予其类人的直觉理解和自适应推理能力。本次报告将从表征（Neural Symbolic Modeling）和推理方式（Learning to inference）两个角度展开，介绍如何能让相机初步具备建模物理动态的能力。
嘉宾简介
vivo影像规划中心研究员。获上海交通大学博士学位（2024），师从杨小康教授、王韫博副教授，研究方向为不完备信息条件下的三维运动推理。当前研究兴趣集中在基于直觉的物理动态推理以及生成式人像编辑。
个人主页
https://syguan96.github.io/
主持人简介
晏轶超
上海交通大学大学
上海交通大学人工智能研究院助理教授，博士生导师。获上海交通大学电子工程系学士、博士学位，法国里昂中央理工学院硕士学位，曾担任阿联酋起源人工智能研究院研究科学家。主要研究方向为计算机视觉、图形学技术及其在虚拟现实、数字多媒体中的应用，发表包括TPAMI、CVPR、NeurIPS在内的国际高水平论文40余篇，先后主持国家自然科学基金青年项目、CCF-阿里巴巴青年科学家基金等项目5项。入选上海市海外高层次人才计划，获2020年度中国图象图形学学会优秀博士论文奖。
长按二维码，关注公众号
GAMES网站: http://games-cn.org
