Online Sparse Scene Coordinates Learning for Real-Time Camera Relocalization

Camera relocalization refers to the problematics of defining the camera pose in known scenes. It is needed in several applications such as augmented reality or robot navigation. However, current camera relocalization systems require an off-line step, that uses the labeled images to construct a scene model. This step takes time, so that is difficult to apply in the augmented reality applications. Thus, we develop an online learning for real-time camera relo-calization system. We present a hybrid method combining machine learning approach and geometric approach for accurate and real-time camera relocalization from each single RGB image independently. For machine learning part, we propose a sparse scene coordinates regression forest, which is able to be fast learned during images capturing. Our system presents how easily user can experience on scenes in desktop-scale.