Yuanwen Yue

This project aims to develop 3D-aware scene representations for AI systems and empower them with semantic understanding and spatial reasoning capabilities. Our focus includes the following questions: (1) How to design scene representations that can efficiently and accurately reconstruct complex 3D environments? (2) How to effectively leverage prior knowledge from 2D foundation models while improving their 3D awareness? (3) How to develop self-supervised learning approaches for visual reasoning on 3D data? Our goal is to design general-purpose models that can operate effectively in the open 3D world, from structured scenes (e.g. indoor and driving scenes) to natural environments (e.g. forests).