Kristen Grauman: Video Understanding for Skill Learning скачать в хорошем качестве

Kristen Grauman: Video Understanding for Skill Learning

Date: March 6, 2025 Speaker: Kristen Grauman, Professor, UT Austin Abstract: The first-person or “egocentric” perspective offers a special window into an agent’s attention, goals, and interactions, making it an exciting avenue for the future of both augmented reality and robot learning. This talk will describe our recent explorations for first-person perception, motivated particularly by learning about human skills from video. In this domain, key challenges are fine-grained activity understanding and relating first- and third- (actor and observer) perspectives. Towards addressing those challenges, we introduce new ideas for view-invariant representations, generating language commentary from visual demonstrations, and navigating massive corpora of how-to videos to discover task structure. I’ll also overview how we are advancing the frontier of egocentric perception for the broader community via large-scale open-sourced datasets called Ego4D and Ego-Exo4D—multi-year, multi-institutional efforts to capture daily-life and skilled activity of people around the world.