Guest Lecture 2: A "Splatacular" Year of 3D Reconstruction
Songyou Peng
Research Scientist at Google DeepMind
April 30, 2025 (Wed), 1:00 p.m. KST
Online (Zoom).
Guest Lecture at CS479: Machine Learning for 3D Data
Minhyuk Sung, KAIST, Spring 2025
Abstract
Accurate, fast 3D reconstruction unleashes the full potential of real-world applications like robotics, AR/VR, and autonomous navigation, yet traditional pipelines remain slow, iterative, and pose-dependent, can not handle dynamic scenes. Over the past year we’ve taken a “Splatacular” path starting with a fully feedforward pipeline that first tackled large-scale, pose-known reconstruction, then learned to infer geometry and camera orientation jointly from just a few unposed views. We then focused on dynamic environments, and merged pose estimation and Gaussian-based scene representation into a single real-time SLAM-style system. This talk will trace that evolution towards truly instant, pose-agnostic, dynamic 3D everywhere.
Bio
Songyou Peng is a Research Scientist at Google DeepMind in San Francisco. He earned his PhD from ETH Zürich and the Max Planck Institute for Intelligent Systems in 2023 under Marc Pollefeys and Andreas Geiger, and then worked as a Senior Researcher/Postdoc at ETH Zürich. Songyou has done internships at Google Research, Meta Reality Labs Research, TUM, and INRIA. His work sits at the nexus of 3D vision and deep learning, most recently driving the development of large-scale foundation models for 3D, from neural implicit representations to multimodal pipelines that unify reconstruction, SLAM, and scene understanding.
-
Image from WildGaussians(https://wild-gaussians.github.io/). ↩