Prof. Dr. Hanno Gottschalk
Technische Universität Berlin

Uncertainty Quantification and Robustness in Recognition of Street Scenes using Neural Networks
(English)

Kolloquiumsvortrag

Donnerstag, 18.04.2024, 10:15 – 11:45 Uhr
Raum N.N.

Gastgeber: Prof. Dr. Emmanuel Müller

Zusammenfassung
Neural networks set new state of the art in the recognition of street scenes, however they are easily fail when confronted with previously unseen out of distribution (OoD) data. OoD can have several meanings - like the same street scene at another time of the day or street scenes where OoD objects occur that belong to a semantic category that is disjoint from those that were present in the the training data. We consider methods that help to identify and localize such OoD objects in semantic segmentation. We discuss, how to robustify neural networks against the domain shift using vision language pre-training. We also consider retrival and unsupervised continual learning of previously unknown objects in street scenes.