Visual Localization - Naver Labs Europe

NAVER LABS Europe is hosting guests from Chalmers University of Technology and the University of Heidelberg who will speak about their respective work on visual localization.

Doors open at 9.45 and the seminars run from 10am to 12pm. Please register online

Date: 15th May 2019


Speaker: Torsten SATTLER, Associate Professor at CHALMERS UNIVERSITY OF TECHNOLOGY in Gothenburg, Sweden

Chalmers university logoAbstract: Visual Localization is the problem of determining the position and orientation from which an image was taken with respect to a known scene. Visual Localization plays an important role in many advanced Computer Vision applications, including autonomous vehicles such as self-driving cars and Augmented / Mixed / Virtual Reality. An especially interesting and hard problem in the context of Visual Localization is that of long-term operation: As the appearance and geometry of the scene changes over time, for example due to illumination or seasonal changes, visual localization algorithms need to be robust to such changes. In the talk, I will present recent work on tackling the long-term Visual Localization problem, ranging from developing benchmarks that can be used to measure the performance of localization methods under different conditions to using semantic scene understanding and machine learning to make visual localization algorithms more robust and reliable.


Speaker: Eric BRACHMANN, Research Associate at Visual Learning Lab at the University Heidelberg in Hannover, Germany

Abstract: Correspondence-based visual localization, as well as other classical computer vision pipelines, often use random sample consensus (RANSAC) for robust optimization of model parameters. In this talk, I discuss different strategies to overcome the non-differentiability of RANSAC which allows to train neural networks in conjunction with RANSAC to directly optimize the desired task loss function. In particular, I discuss how to predict image-scene correspondences while allowing for outlier predictions, how to predict correspondence sampling weights for guided RANSAC hypotheses search, and how to integrate RANSAC in a mixture of experts model.

Visual Learning Lab Logo

Ceci correspond à une petite biographie d'environ 200 caractéres