Abstract: In this talk, I will present our recent work at Adobe Research on 3D reconstruction and generation using neural models. The talk will first discuss our work on neural field representations, which allow for fast, compact, and realistic 3D reconstruction and rendering. Then, we will investigate how to estimate neural fields efficiently with deep networks in a feed-forward manner to enable fast reconstruction. We will also look into the strong connections between 3D reconstruction and generation and discuss how reconstruction can serve as a foundation for generation tasks. Our latest research utilizes large transformer models to enable fast and realistic 3D reconstruction and generation on various tasks, ranging from multi-view reconstruction, to single-view reconstruction and text-to-3D generation.
Bio: Zexiang Xu is currently a research scientist at Adobe Research. He received his Ph.D. from UC San Diego in 2020, where he was advised by Prof. Ravi Ramamoorthi. His research focuses on 3D computer vision and graphics, covering neural rendering, inverse rendering, 3D representation, 3D reconstruction, and 3D generation. (https://cseweb.ucsd.edu/~zex014/)
