We look forward to seeing you online on Thursday, 9/21.
Abstract: Maps help us navigate the world. For centuries, humans relocalised themselves using birds-eye 2D maps. Today, computer vision enables precise 6DoF relocalisation of a camera’s position and orientation within 3D maps pre-built from sets of images - a task referred to as visual relocalisation. An established recipe for visual relocalisation is to build explicit maps consisting of a structure-from-motion point cloud, and solve for a camera’s pose using feature matching. About a decade ago, with the advent of machine learning in vision, researchers explored the possibility of encoding implicit maps in learned models. The talk will discuss the main approaches towards implicit, learned maps for visual relocalisation: Scene coordinate regression, absolute pose regression and relative pose regression. The talk will culminate in the question of when does a visual relocalisation representation become so minimal that it essentially can be called “map-free”.
Bio: Eric Brachmann is a staff scientist at Niantic. He received a doctorate in 2018 by the TU Dresden (Germany), and he was a post-doctoral researcher at the Visual Learning Lab of Prof. Rother at the University Heidelberg, until 2020. He works on object pose estimation [ICCV’15, CVPR’16,’17, ECCV’14,’18], camera re-localization [CVPR’16,’17,’18,’23, ICCV’19,’21, 3DV’21, TPAMI’21, ECCV’22], discrete feature matching [ICCV’19, CVPR’20, ECCV’22], and robust estimation [CVPR’17,’21,’23, ICCV’19]. He is an expert in scene coordinate regression, which is a core element in state-of-the-art learning-based localization techniques. He publishes his work at the leading computer vision conferences (ECCV, ICCV, CVPR), is an active reviewer (CVPR, ICCV, ECCV, NeurIPS, T-PAMI, IJCV, JMLR) and area chair for WACV’23. He co-organized four tutorials on visual localization [ECCV’18, ICCV’19,’21, CVPR’23], and four workshops on 6D pose estimation of objects [ICCV’19,’23,ECCV’20,’22].