课程题目：Optimization Time Integration for Solids and Fluids
授课讲者：Minchen Li (李旻辰), Carnegie Mellon University
课程摘要：Second-order optimization methods, such as Newton's
Method, are critical not only in geometry processing for applications like shape deformation and mesh
parameterization but also in the robust and accurate simulation of solid and fluid dynamics. This talk
provides a high-level overview of optimization time integration methods, starting from a geometric
perspective focused on distortion minimization. Participants will learn how to extend distortion
minimization methods to an elastodynamic simulation framework and will explore methods for simulating a
variety of materials and phenomena, including cloth, hair, stiff objects, contacts, and fluids. The session
also links to a comprehensive online book and a set of illustrative Python examples for the elastodynamic
contact part to enhance understanding. By the end of this course, attendees will gain a deeper insight into
the close connection between geometry processing and physics-based simulation.
讲者简介：Minchen is an assistant professor in the Computer
Science Department at Carnegie Mellon University, having joined in September 2023 after leaving his role as
an assistant adjunct professor at UCLA Department of Mathematics. He received Ph.D. from the University of
Pennsylvania, advised by Chenfanfu Jiang. Minchen is a winner of the 2021 ACM SIGGRAPH Outstanding Doctoral
Dissertation Award for the development of the Incremental Potential Contact (IPC) method. His current
research focuses on integrating physics-based simulation with AI for computer graphics, visual computing,
robotics, and computational mechanics.
课程摘要：Numerical simulation of physical systems has become
an increasingly important scientific tool supporting various research fields. Despite its remarkable
success, simulating intricate physical systems typically requires advanced domain-specific knowledge,
meticulous implementation, and enormous computational resources. With the surge of deep learning in the last
decade, there has been a growing interest in the machine-learning and graphics communities to address these
limitations of numerical simulation with deep learning. This course provides a gentle introduction to this
topic for audiences interested in exploring this trend but with little to modest machine-learning or
physics-simulation backgrounds. We begin with a brief overview of the numerical simulation framework on
which we ground our discussion of deep-learning methods. Next, the course provides a possible classification
of several hybrid simulation strategies based on the roles of learning and physics insights incorporated. We
then review the implications of such deep-learning strategies and discuss some practical considerations in
combining deep learning and physics simulation. Finally, we briefly mention several advanced
machine-learning techniques for further exploration.
讲者简介：
杜韬，清华大学交叉信息研究院助理教授，博士生导师。杜韬博士毕业于麻省理工学院计算机图形学实验室，主要研究方向为图形学中的物理仿真与计算设计。他的相关工作主要发表在计算机图形学和机器学习领域顶级期刊和会议（ACM
TOG, SIGGRAPH North America/Asia, ICLR, ICML, NeurIPS)上，并受到多家知名科技媒体（WIRED，MIT News, IEEE Spectrum,
TechCrunch等）的关注与报道。此外，他多次担任SIGGRAPH North America/Asia技术论文程序委员会委员并多次获评NeurIPS/ICML优秀审稿人。
课程摘要：图形的真实感渲染离不开基于Monte
Carlo估计进行复杂系统和高维积分求解，其核心问题之一就是进行有效和高效的路径采样以降低样本的方差、加速计算收敛，而这也是影响图形渲染的质量和效率的关键。本报告将围绕近期开展的图形真实感渲染相关研究工作和发表在ACM
TOG (Siggraph)和IEEE
TVCG的一系列进展展开，介绍在图形渲染中光子映射的样本统计检验处理方法；双向路径追踪BDPT中的路径采样和概率连接方法、多重重要性采样和重采样方法；进一步介绍基于神经网络的路径引导采样方法等，并讨论基于神经网络方法的多种可行性。
课程摘要：
近年来以NeRF、3DGS为代表的神经表达技术为三维场景的重建与渲染提供了新的技术途径，在重建稳定性、渲染真实感等方面相较于传统MVS方法也体现了巨大优势，但在处理大规模场景和动态场景等方面仍存在局限。本次报告将介绍我们课题组在这些方面的最新工作，主要包括大规模场景的高效表达方法Level
of Gaussians，动态街景的建模方法StreetGaussians，以及动态场景神经表达重建与渲染开源软件EasyVolcap。
课程摘要：Obtaining compact building models has significant
social relevance as it can enable better urban planning, design, and management. Accurate and detailed
digital 3D city models can help mitigate the negative impacts of urbanization, such as traffic congestion,
pollution, and inefficient land use. Though advances in laser scanning and 3D computer vision have enabled
efficient and effective data acquisition of urban environments, obtaining faithful 3D surface models of
urban buildings remains an open challenge. In this talk, I will share my experiences from the past years in
the 3D reconstruction of urban buildings. I will present a series of algorithms for reconstructing simple
polygonal surface models for piecewise planar objects and several extensions, including deep learning based
methods, for reconstructing large-scale urban buildings. Finally, we will discuss trends and future
development ideas in this field.
讲者简介：Liangliang Nan received his B.S. degree in material
science and engineering from Nanjing University of Aeronautics and Astronautics (NUAA), China, in 2003 and a
Ph.D. degree in mechatronics engineering from the Graduate University of the Chinese Academy of Sciences,
China, in 2009. After obtaining his Ph.D. degree, he worked as an assistant researcher (in 2009) and then as
an associate researcher (in 2011) at the Shenzhen Institute of Advanced Technology (SIAT), Chinese Academy
of Sciences. From 2013 to 2018, he was a research scientist with the Visual Computing Center at King
Abdullah University of Science and Technology (KAUST) in Saudi Arabia. He is currently an associate
professor with the Faculty of Architecture and the Built Environment at Delft University of Technology (TU
Delft) in the Netherlands. At TU Delft, he leads the AI lab on 3D Urban Understanding (3DUU). His research
interests include computer graphics, computer vision, machine learning, and 3D geoinformation.