|
|
Visual Localization in Natural Environments
Brejcha, Jan ; Sattler, Torsten (oponent) ; Matas, Jiří (oponent) ; Čadík, Martin (vedoucí práce)
We focus our work on camera position and orientation estimation given a query photograph; we call this problem visual geo-localization. Specifically, we focus on photographs captured in natural, mountainous environments. We introduce a thorough review of state-of-the-art computer vision methods, datasets, and evaluation practices for visual geo-localization problems. The survey revealed that researchers usually cast visual geo-localization in natural environments as a similarity or a correspondence search between an input photograph and a terrain model; we call this problem the cross-domain matching. We identified three main goals to improve over the state of the art in visual geo-localization in mountainous environments using cross-domain matching: (I) the need for new datasets for training, validation, and evaluation of cross-domain visual geo-localization algorithms, (II) the need to verify whether the cross-domain matching algorithms may benefit from using different features-horizon lines, edge maps, semantic segmentation, and satellite imagery, (III) the need to illustrate the usefulness of visual geo-localization methods by developing novel applications. In this thesis, we thoroughly describe our research studies to illustrate how we examined particular goals. We introduce several novel datasets for evaluation and training of cross-domain matching methods. These novel datasets allowed us to propose a novel method for cross-domain photo-to-terrain matching using a combination of semantic segments and classic edge-based features. We illustrate the benefits of our novel approach over the state of the art on camera orientation estimation. Furthermore, we propose a meta-algorithm based on a cross-domain Structure from Motion for a weakly supervised acquisition of cameras aligned with the synthetic terrain. This novel cross-domain data acquisition scheme allowed us to train a compact cross-domain keypoint descriptor. We illustrate the descriptor performance by estimating full camera pose by matching the query photograph to the rendered terrain model. Finally, we demonstrate a practical usability of outdoor visual geo-localization by designing a novel application of photography presentation on a computer screen or in virtual reality. Moreover, we illustrate that our novel presentation method helps the user with complex outdoor scene understanding and improves self-localization in unvisited outdoor environments.
|
|
|
Visual Localization in Natural Environments
Brejcha, Jan ; Sattler, Torsten (oponent) ; Matas, Jiří (oponent) ; Čadík, Martin (vedoucí práce)
We focus our work on camera position and orientation estimation given a query photograph; we call this problem visual geo-localization. Specifically, we focus on photographs captured in natural, mountainous environments. We introduce a thorough review of state-of-the-art computer vision methods, datasets, and evaluation practices for visual geo-localization problems. The survey revealed that researchers usually cast visual geo-localization in natural environments as a similarity or a correspondence search between an input photograph and a terrain model; we call this problem the cross-domain matching. We identified three main goals to improve over the state of the art in visual geo-localization in mountainous environments using cross-domain matching: (I) the need for new datasets for training, validation, and evaluation of cross-domain visual geo-localization algorithms, (II) the need to verify whether the cross-domain matching algorithms may benefit from using different features-horizon lines, edge maps, semantic segmentation, and satellite imagery, (III) the need to illustrate the usefulness of visual geo-localization methods by developing novel applications. In this thesis, we thoroughly describe our research studies to illustrate how we examined particular goals. We introduce several novel datasets for evaluation and training of cross-domain matching methods. These novel datasets allowed us to propose a novel method for cross-domain photo-to-terrain matching using a combination of semantic segments and classic edge-based features. We illustrate the benefits of our novel approach over the state of the art on camera orientation estimation. Furthermore, we propose a meta-algorithm based on a cross-domain Structure from Motion for a weakly supervised acquisition of cameras aligned with the synthetic terrain. This novel cross-domain data acquisition scheme allowed us to train a compact cross-domain keypoint descriptor. We illustrate the descriptor performance by estimating full camera pose by matching the query photograph to the rendered terrain model. Finally, we demonstrate a practical usability of outdoor visual geo-localization by designing a novel application of photography presentation on a computer screen or in virtual reality. Moreover, we illustrate that our novel presentation method helps the user with complex outdoor scene understanding and improves self-localization in unvisited outdoor environments.
|
host ::
RSS.