Spatial Scene Understanding with Neural Scene Representations
COM2 Level 4
Executive Classroom, COM2-04-02
Abstract:
Novel scene representations like neural radiance fields and Gaussian splatting have enabled disruptive progress in the field of offline 3D reconstruction. This talk will demonstrate that these advances can be well transferred to the online setting in which scene reconstruction and camera tracking are jointly computed in a continual manner. A key ingredient for this success is differentiable rendering that allows for both novel view synthesis and effective self-supervised learning of scene representations like neural radiance fields. Ideally, this technology can be a steppingstone for the spatial-temporal reasoning of future robots or autonomous cars.
Brief Bio:
Martin R. Oswald is an Assistant Professor in the Computer Vision Group at the University of Amsterdam and scientific lab manager of ATLAS lab, a collaboration with TomTom. Before he was a senior researcher, lecturer and deputy lab director in the Computer Vision and Geometry lab at ETH Zurich, headed by Prof. Marc Pollefeys. He obtained his PhD in Computer Vision from TU Munich under the supervision of Prof. Daniel Cremers in 2015. Previously, he received a masters degree in civil engineering from the University of Technology Federico Santa Maria in Valparaiso, Chile in 2008 supported by a DAAD fellowship program. Before he studied computer science at TU Dresden and the University of Technology Sydney and obtained a master degree (Diplom) from TU Dresden in 2007. His works on 3D reconstruction received several awards including best paper honorable mention awards at 3DV 2024 & 2019, the ACCV Honorable Mention Award in 2010 and the DAGM 2009 Paper Award.