Pose Estimation From Camera Images for Underwater Inspection

High-accuracy localization is essential for underwater reinspection missions, which often require revisiting sites with complex structures for inspection and maintenance. Traditional localization systems—such as inertial navigation, Doppler velocity logs, and acoustic positioning—frequently fall short in accuracy or cost-effectiveness for these tasks. Machine learning-based visual relocalization presents a promising alternative, estimating poses from monocular images captured by onboard cameras using models trained on data from previous deployments. In this work, we evaluate the performance of such learning-based estimators in both clear and turbid water environments, examining the effects of color information, model architecture, and training data diversity. We further propose a novel view synthesis-based strategy to augment training data, enhancing pose estimation at previously unseen viewpoints. Finally, we improve localization robustness by fusing pose estimates with additional sensor inputs via an extended Kalman filter, resulting in smoother and more accurate trajectories.