“Large Reconstruction Model for Efficient 3D Reconstruction and Generation”
ABSTRACT
Over the past year, the large language model has achieved significant milestones, approaching humanlike intelligence across various domains. However, there has been limited investigation into largescale 3D reconstruction in the literature. In this talk, I will primarily focus on our recent advancements in largescale 3D reconstruction.
I will start with an introduction to the basics of the Largescale Reconstruction Model (LRM), aiming to develop a robust and highly generalizable 3D reconstruction system utilizing highquality 3D data. I will also explain how LRM can be used to efficiently perform highquality textto3D and imageto3D generation tasks, such as Instant3D and DMV3D.Finally, I will highlight our recent work, specifically our progress in largescale 3D reconstruction using Gaussian Splatting (GRM). As a largescale reconstructor, GRM can reconstruct a 3D asset from sparseview images in about 0.1 seconds. Moreover, GRM shows promising potential in generative tasks, such as textto3D and imageto3D, through its integration with existing multiview diffusion models.
PRESENTER
https://justimyhxu.github.io/
Yinghao Xu is a postdoctoral researcher at the Stanford Computational Imaging Lab, Stanford University, where he works under the guidance of Professor Gordon Wetzstein. Previously, he was a Ph.D. student at the Multimedia Lab (MMLab) at The Chinese University of Hong Kong. He has a deep interest in the intersection of Computer Graphics and Computer Vision. His current research focuses on generative models and neural rendering, particularly in the area of 3D generative models. Many of his papers have been accepted for oral presentation at top conferences such as CVPR, ECCV, and ICLR. One of his papers was nominated as a best paper candidate at CVPR 2020.