세미나안내
From Scenes to Individuals: A Long-term Agent-centric Foundation Model for Human Behavior
- 등록일2026.10.29
- 조회수24
-
세미나 일정2026.10.30 금
-
연사박현수(Playtag 대표)
[Abstract]
Today's video foundation models understand scenes. They can describe what happens in a clip, but they lose track of who did it once that person leaves the frame. Yet most questions we ask about human behavior are about specific individuals over long periods: how a child's social play changes over a semester, or whether a worker's movement drifts toward risk over weeks. In this talk, I will introduce LAM (Long-term Agent-centric Model), a foundation model that moves video understanding from an Eulerian, scene-centric view to a Lagrangian, agent-centric one. LAM couples persistent multi-person tracking with multi-level attention over the scene, proxemics, and the person, so that identity and behavior are learned jointly rather than stitched together afterward.
I will argue that this shift requires rethinking data as much as architecture. Long-term identity calls for multi-view "5D" capture, unscripted behavior of many people across many places, and consent from those being recorded. This work builds on our earlier research in multiview geometry and self-supervised social signals. I will share lessons from training on millions of hours of real-world video and from deploying the model in early education, industrial safety, and sports. I will also show where current video-language models break down, for example by inventing identities or accepting false premises. I will close with open problems: grounding language in individuals, measuring object state change for robot learning, and making such models cheap enough to run everywhere.
[Biography]
Hyun Soo Park is the founder and CEO of Playtag, an AI company building foundation models for long-term human behavior understanding. He was previously an Associate Professor of Computer Science and Engineering at the University of Minnesota, and before that a Postdoctoral Fellow in the GRASP Lab at the University of Pennsylvania. He received his Ph.D. from Carnegie Mellon University. His research centers on computer vision for behavioral imaging. He has received the NSF CRII and NSF CAREER Awards and a CVPR 2021 Best Paper Honorable Mention, and serves as an Area Chair for ECCV.
