3D point cloud simplification for image-based localization

Recent advances in image-based localization have made it possible to estimate the camera pose relative to 3D point clouds. In a multimedia cloud computing paradigm, large-scale 3D point clouds can be stored on the server side, and this enables us to implement image-based localization applications on mobile devices toward many cloud-media applications and services. When scaling the 3D point cloud database up to a world-wide level, the prohibitive memory footprint of such large-scale point clouds is a heavy burden to the server side. In this paper, we present a method to compute a compact subset of a 3D point cloud for image-based localization while achieving acceptable localization performance. In our method, we consider both coverage of database images and emphasis on 3D points of high spatial significance, and we show that our method outperforms the state-of-the-art methods on three public datasets especially when the size of the resultant subset is small. In essence, the priority of system resource is given to points of high spatial significance.