Abstract

In computer vision applications such as mobile robotics and autonomous driving, 3D scene understanding refers to the problem of inferring both the semantics and the geometry of a scene from sensor data. With the deep learning revolution, massive improvements have been made in semantic tasks – i.e. answering what we are looking at. The impact of deep learning on geometric tasks – i.e. answering where things are and what shape they have – has, however, not yet been as prominent. In this talk, some recent work on deep learning for geometric tasks will be covered. One example will be taken from work on 3D object detection carried out by Eskil and his colleagues, along with some discussions on how thinking about uncertainty and data representations can be helpful when designing machine learning models.

Eskil Jörgensen

Research Scientist @ Zenuity

Eskil was trained in engineering physics and wireless communication at Linköping University, National Taiwan University, and Ericsson before moving to Gothenburg to dive deep into the world of AI. He is now pursuing an industrial PhD in deep learning and computer vision at Zenuity and Chalmers, with a focus on 3D scene understanding for automotive applications. The one question occupying Eskil's mind is: how far can you push the limits of camera-only perception systems?