Viktor Larsson, LTH: Mapping and Localization for AR

For augmented reality devices to persist content across sessions it is necessary to both build maps of the environment and to localize within them. The current benchmarks for visual localization are focused on single-image localization. However, in the context of AR, more information is usually available, e.g. temporal data (image sequences), IMU or radio signals. In this talk I will present our recent work on benchmarking localization and mapping in the context of AR. We have introduced a new dataset which contain diverse and large-scale scenes recorded with head-mounted and hand-held AR devices, together with an accurate ground truth obtained via laser scanners. While we used the dataset for benchmarking, we believe it can be useful in many other contexts as well.

Viktor Larsson is an assistant professor in the Mathematical Imaging Group at Lund University. He previously worked as a PostDoc and senior researcher at the Computer Vision and Geometry group lead by Marc Pollefeys at ETH Zurich. His research interests include robust estimation problems that appear in 3D computer vision (e.g. Structure-from-Motion, visual localization and SLAM). He is currently exploring how to best integrate deep learning with classical geometry-based pipelines. In particular, identifying and replacing hand-crafted heuristics or implicit assumptions with more data-driven alternatives.