Modern applications of visual localization requires carefully balancing the trade-off between pose estimation accuracy, computational efficiency and scalability. However, existing methods for visual localization struggle to perform accurate, real-time localization, especially in large scale environments.

To address this problem, this thesis presents AnchorLoc, a framework which leverages real-time object detection to detect visual anchors and optimize the visual localization pipeline to dramatically improve efficiency and scalability. AnchorLoc automatically extracts stable and distinct visual anchors from image sequences captured in the environment. These visual anchors are then detected from the query image and used to optimize the search space during image retrieval and local featu...[ Read more ]

Modern applications of visual localization requires carefully balancing the trade-off between pose estimation accuracy, computational efficiency and scalability. However, existing methods for visual localization struggle to perform accurate, real-time localization, especially in large scale environments.

To address this problem, this thesis presents AnchorLoc, a framework which leverages real-time object detection to detect visual anchors and optimize the visual localization pipeline to dramatically improve efficiency and scalability. AnchorLoc automatically extracts stable and distinct visual anchors from image sequences captured in the environment. These visual anchors are then detected from the query image and used to optimize the search space during image retrieval and local feature matching processes. Additionally, we collect a large scale visual localization dataset comprised of image sequences and 3D reconstruction of an area within a university campus.

Experimental results reveal that AnchorLoc reduces localization time by 83% on our campus dataset and by 69% on the Cambridge Landmarks dataset without significantly degrading the pose estimation accuracy compared to HLoc, an accurate hierarchical localization method. AnchorLoc is also more accurate and faster than SLD, a localization method which takes a similar approach at the keypoint level. Lastly, this thesis suggests research directions for future development of anchor-based methods for more robust and efficient visual localization.