Dr. Lior Yariv (Stanford) : “Learning Signed Distance Representations for 3D Modeling”

Speaker: Dr. Lior Yariv (Stanford)

Title: Learning Signed Distance Representations for 3D Modeling

Video: 

Abstract: A central question in learning 3D scenes is choosing an appropriate and efficient scene representation, one that faithfully represents the scene’s underlying geometry.
The goal of this talk is to present Signed Distance Functions (SDFs) as a core representation for 3D modeling. SDFs offer a flexible and expressive way to model a wide variety of closed surfaces with complex topology, while also allowing for straightforward surface extraction. Moreover, the geometric information encoded in SDFs facilitates accurate rendering and supports favorable reconstruction.
We present the potential of the SDF representation in addressing key problems in computer vision and computer graphics, focusing on two fundamental tasks: multi-view surface reconstruction and 3D generation. For multi-view surface reconstruction, we introduce novel methods for learning SDF-based representations directly from 2D images, by integrating SDFs into differentiable volume rendering pipelines. Towards 3D generation, we present a scalable and expressive surface representation tailored for training flow-based generative models on large-scale 3D shape datasets. Together, these contributions demonstrate the versatility and effectiveness of SDFs in addressing key challenges in learning meaningful 3D representations under different supervision settings.

Bio: Lior is a Postdoctoral Scholar at Stanford University working with Prof. Gordon Wetzstein. She received her doctorate from the Department of Computer Science and Applied Mathematics at the Weizmann Institute of Science under the supervision of Prof. Yaron Lipman. Her research focuses on the intersection of 3D geometry and generative modeling. She is particularly interested in devising spatially-aware methodologies that improve how we understand and synthesize 3D environments.