Learning Visual Foundation Models for Perceiving 3D Humans

The world is entering the age of autonomy, where systems like self-driving cars and assistive robots face the challenge of perceiving and understanding human behavior in complex, dynamic environments through sensory inputs like images. While human motion and behavior have been studied extensively using visual data, existing methods often struggle in open-world settings due to their inability to generalize beyond domain-specific training data. The goal of this project is to learn generalizable visual representations from large-scale training data and create a foundational model that supports various human analysis and modeling tasks, grounded in 3D environments, in an open-world setting. Our key objectives include (1) Devising new human-centric representations and algorithms for accurate 3D human motion and appearance reconstruction from in-the-wild videos. (2) Developing advanced generative models synthesizes large-scale, realistic virtual humans, emphasizing controllability and efficiency. (3) Designing physics-guided methods for long-duration human-scene interaction capture using embodied devices. (4) Studying efficient learning algorithms and unified model architectures for 3D human foundation models that enable efficient fine-tuning and generalization across various human-centric computer vision tasks. We aim to establish the computational and algorithmic groundwork for human-centric visual foundation models, with advancement in generative modeling, human-centric representation learning, and visual data simulation. Our work will facilitate the development of autonomous AI agents for various sectors, including autonomous driving, extended reality, education, and manufacturing. Ultimately, this project is committed to advancing human-centric AI technologies, fostering more robust and effective human-AI interaction, and paving the way for practical applications that significantly improve everyday life and industrial processes.