Abstract: Robust and accurate self-localization and real-time tracking of egomotion are key challenges for augmented reality and autonomous machines. In this talk, I will present our recent works in visual localization and odometry which address these challenges in the context of mobile devices and machines. The first part of the talk presents our probabilistic inertial-visual odometry approach for robust egomotion estimation. The method is based on an information fusion framework and employs low-cost IMU sensors and rolling shutter cameras in a standard smartphone. We demonstrate results computed in real-time on an Android phone and provide comparisons against proprietary solutions. The second part of the talk presents our recently published hierarchical scene coordinate classification and regression approach for visual localization. The approach is based on a hierarchical neural network which is trained to regress the mapping between raw image pixels and 3D coordinates in the scene. Thus, the matching between a query image and a pre-built 3D model is implicitly performed by the forward pass through the network. The proposed hierarchical network outperforms the baseline regression-only network and allows us to train compact models which scale robustly to large environments. It sets a new state-of-the-art for single-image RGB localization performance on the 7-Scenes, 12-Scenes, Cambridge Landmarks datasets, and three combined scenes.

Speakers:  Juho Kannala

Affiliation: Department of Computer Science, Aalto University.

Place of Seminar:  Zoom (Available now on Youtube)