Computer Vision News - November 2022

28 Congrats, Doctor! In construction, delivery, transportation, care, or household work there are many applications where robots could take over dangerous tasks, help to reduce energy and resource use, or enable novel design principles. However, they require robots to operate in our everyday environments instead of factory floors, which often requires semantic scene understanding. Data-driven algorithms, especially deep learning, have greatly improved the capabilities of machines to detect and identify objects. Yet, these methods fail under domain shift and in the presence of unseen object types. Hermann’s thesis investigates the problemof robotic scene understanding in open-world, everyday environments. The proposed systems are able to identify unknown parts of a scene, and even adapt and improve their perception capabilities in these environments fully autonomously. In a first step, his thesis introduces a benchmark (fishyscapes.com ) to measure how well robotic perception methods can identify outliers. It focuses on anomalies for semantic segmentation in urban driving, where unknown categories should be correctly segmented from images. As part of a larger collaboration a second benchmark was later added (segmentmeifyoucan.com) that focuses on real-world images and detection of any obstacle on the road. The analysis of existing solutions then revealed that good anomaly segmentation methods often had lower segmentation accuracy on the known Hermann Blum recently completed his PhD in the Autonomous Systems Lab of ETH Zürich. His research focuses on enabling robots to semantically understand their environments in complex open-world settings. Probably best known for the fishyscapes. com benchmark, he first worked on anomaly detection in semantic segmentation and later on self- improving perception systems. He recently joined the Computer Vision and Geometry Lab of ETH Zürich. Congrats, Doctor Hermann!