Camera Pose Refinement for City-Scale  
Visual Positioning Systems  

Given a low level of detail city model and the initial camera pose estimate of an image, the task is to find a more accurate pose. We develop a simple yet effective pose refinement approach utilizing image features and rendering based on intial pose. The approach is applied to large scale visual positioning systems and signficiantly improves localization accuracy in a multi-city benchmarking dataset. Several examples are presented below, showing that model rendering and images are much better aligned after our pose refinement approach.


Please address questions to