6-DoF Pose Refinement via Sparse-to-Dense Feature-Metric Optimization

Abstract

In this report, we detail our submission to the CVPR 2020 visual localization challenge. Previous work on sparse-to-dense matching showed outstanding robustness to extreme conditions, but lacks precision due to the low resolution of deep feature maps. We introduce a simple algorithm to refine the estimated pose based on the feature-metric error, and demonstrate improved localization accuracy. This, combined with better feature selection, results in state-of-art night localization on the RobotCar dataset.