【講演概要】3D reconstruction and generation are the key technology of real-world digitalization, which has important applications in intelligent manufacturing, robotics, metaverse and other fields. Among all the 3D representations, Neural Radiance Field (NeRF) and 3D Gaussian Splatting (3DGS) stand out as prominent techniques. NeRF, a deep learning-based approach, excels at generating realistic 3D images by learning scene geometry and lighting. It requires massive and high-quality data, and struggles for dynamic scenes. In contrast, 3DGS, with its explicit representation, offers efficient rendering and ease of editing through 3D Gaussian functions defined by position, opacity, covariance, and spherical harmonics. In this talk, I will introduce our recent work on 3D reconstruction and generation. These techniques enhance NeRF’s adaptability and efficiency for complex scenes, dynamic scenes, and sparse and noisy input poses. They also enable more flexible and controllable 3D editing and generation with 3DGS. Collectively, these contributions advance the field of 3D scene generation and reconstruction, enabling more accurate and efficient creation and manipulation of 3D content.
【略歴】Shi-Min Hu is currently a professor in the Department of Computer Science and Technology, Tsinghua University. He received his Ph.D. degree from Zhejiang University in 1996. His research interests include digital geometry processing, video processing, rendering, computer animation, computer aided geometric design and deep learning. He has published more than 100 papers in journals and refereed conferences. He is the Editor-in-Chief of Computational Visual Media , and on the editorial boards of several journals, including Computer Aided Design and Computer & Graphics. Professor Hu is a member of Chinese Academy of Science, an IEEE Fellow and winner of CCF Wang Xuan Award.