151-0652-00 MaP Robotics, Vision, and Controls Talks

Abstract

Introduction day with an overview of the Robotics, Vision, and Controls Talks series, including format, goals, and this semester's themes, followed by a talk by one of the organizers.

Abstract

In this talk, I discuss long-term point tracking from a robotics perspective, where models operate online, in real time, and under strict memory constraints. I begin with how visual foundation models improve robustness and viewpoint invariance, then introduce a causal transformer formulation where points are queries and frames are processed sequentially. I focus on temporal propagation with compact memory for long-term consistency, leading to Track-On, and its extension Track-On2 for improved efficiency and real-time performance. Finally, I cover verifier-guided self-training for adapting synthetic-trained trackers to unlabeled real-world videos.

Bio

Fatma Guney is an Assistant Professor at Koc University. She received her PhD in 2017 from the Max Planck Institute in Tubingen. Her research focuses on autonomous driving and 3D vision, with interests in geometry, motion, and uncertainty.

Bio

Antonino Furnari is an Associate Professor at the University of Catania and a member of the Image Processing Laboratory. His research focuses on embodied egocentric vision for understanding and anticipating human actions to enable assistive wearable technologies.