CNN Models for Accurate Pose Localization

Visual Localization is the problem of estimating the camera pose of a given image relative to a visual representation of a known scene. The estimation process becomes increasingly difficult due to factors such as a large quantity of outliers and changes in viewpoints. To overcome these challenges, we leverage semantic segmentation algorithms to enhance the localization performance. We propose three distinct approaches that effectively combine semantic and visual information, allowing us to assign more than one label to each keypoint. We compare the performance of these methods against state-of-the-art techniques and evaluate the scenario where only mono labels are utilized. The evaluation is conducted on four datasets (Dubrovnik, Rome, Aachen, and Vienna). Through extensive experimentation, we demonstrate that by incorporating both visual and semantic information, the accuracy of pose estimation can be significantly improved in terms of both time efficiency and precision.