Visual Geo-Localization and Camera Pose Estimation in Natural Environments

This thesis presents my work, along with the contributions of my research group CPhoto@FIT, on the challenging task of visual geo-localization and camera pose estimation in natural environments. The primary objective is to determine the geographical position and orientation of a camera using visual information from captured images. Natural environments, characterized by irregular landscapes, self-similar features, and sparse datasets, pose unique challenges compared to more structured settings, such as urban areas. We developed novel methods for camera orientation and pose estimation, utilizing techniques such as silhouette edge matching, semantic cues, line correspondences, and cascaded attention mechanisms. To support training and evaluation in outdoor localization tasks, we introduced three datasets—GeoPose3K, Alps100K, and CrossLocate. These datasets provide a diverse range of visual and synthetic modalities, including semantic segmentation, depth maps, and silhouette maps.
Additionally, we contributed to skyline (horizon line) detection, a critical component of outdoor geo-localization techniques. We further contributed by proposing two complementary approaches to image geo-localization in natural environments. The first method is specifically designed for scenarios with extensive coverage of community-contributed photographs, ensuring high localization accuracy in such regions. In contrast, the second approach, which relies on image retrieval from a synthetic database, offers a
groundbreaking solution by enabling global operation, irrespective of the availability of community photographs. Finally, we explored applications of visual geo-localization in areas such as augmented reality, automatic label placement, and depth map estimation. The methods presented in this work achieve state-of-the-art performance across various benchmarks and real-world scenarios.