Exploiting Spatial and Co-visibility Relations for Image-Based Localization

Abstract

Image-based localization techniques aim to estimate the position and orientation from which a given query images was taken with respect to a 3D model of the scene. Recent advances in Structure-from-Motion, which allow us to reconstruct large scenes in little time, create a need for image-based localization approaches that handle large-scale models consisting of millions of 3D points both efficiently and effectively in order to localize as many query images as possible in as little time as possible. While multiple efficient localization methods based on prioritized feature matching have been proposed recently, they lack the effectiveness of slower approaches. In this chapter, we show that we can increase the effectiveness of approaches based on prioritized 2D-to-3D matching at little to no additional run-time costs by exploiting both spatial and co-visibility relations between the 3D points in the model. The resulting localization framework incorporates both 2D-to-3D and 3D-to-2D matching and achieves state-of-the-art efficiency and effectiveness.