Decision Forest For Efficient and Robust Camera Relocalization

To robustly estimate the pose, classical methods assume some geometrical and temporal assumptions (SfM: Structure from Motion, SLAM: Simultaneous Localization and mapping). These approaches take a pair of images as input and establish correspondences based on global strategy (using the whole image information) or sparse strategy (using key-points features). These correspondences allow solving a set of linear equations related to the 3D information and camera pose in that environment. These past years, machine learning has been considered as an efficient way to tackle different problems in image processing and computer vision fields. To handle the task in hand, we propose to learn directly the mapping function between the acquired information from the camera and its pose using sparse decision forest. We achieved state-of the-art results on public and on our databases.