ELLIS Delft Talk by Holger Ceasar: Building machine perception systems with (almost) no labels
06 februari 2024 16:00 t/m 17:00 - Locatie: Hybrid: Building 28, Room Hilbert (2.W510) / Zoom - Door: ELLIS Delft | Zet in mijn agenda
by Holger Caesar | Delft University of Technology
Abstract
This talk explores the evolution and future direction of perception systems for autonomous vehicles. I argue that the first generation of perception systems consisted of manually labeled datasets (nuScenes) and object detectors (PointPillars) trained with full supervision. The second generation seeks to automatically label vast amounts of driving data (nuPlan). However, the underlying offline perception system is still trained with thousands of human annotated sequences, which is costly and inflexible. I envision a third generation that requires little to no human labeling effort and uses techniques like active learning, object discovery, self supervised learning and foundation models. These methods reduce human annotation efforts by several orders of magnitude and allow us to flexibly adapt to new robotics applications, such as bikes, trains and boats.