Vabljeni na predavanje dr. Aljoša Ošepa, ki bo v petek 3. 2. 2023 ob 14.15 na FRI v P21. Ajoša Ošep je doktoriral na Univerzi RWTH Aachen, delal kot podoktorski raziskovalec na TUM, trenutno pa je raziskovalec na Robotics Institute, Carnegie Mellon University v Pittsburghu. Njegovo raziskovalno področje obsega računalniški vid, robotiko in strojno učenje. Predavanje bo v angleščini.

Naslov: Learning to Understand The World From Video

Povzetek: The future of AI is embodied. Think of embodied Alexa that can assist humans in their daily lives or robo-taxis that can transport people safely and efficiently. However, achieving this future requires solving several technical challenges, particularly in the realm of perception. In particular, we need to answer questions such as where I am or where I need to go (localization), what is around me (detection), and how it moves (tracking and forecasting). As a community, we have made substantial progress in these research fields, mostly relying on supervised learning. However, the data labeling is costly, and the “tail” of the object class distribution is unbounded. Therefore, we need a different set of learning and modeling tools to continue progressing and reach the maturity level needed for the deployment at scale. This motivates the general research direction I will discuss in my talk: learning to detect, track, segment objects, and forecast their future trajectories by simply watching a video or, more broadly speaking, via self-supervised learning from streams of sensory data, such as video or lidar sequences.