|
|
Visual Camera Orientation Estimation using Machine Learning
Kubička, Martin ; Polášek, Tomáš (oponent) ; Čadík, Martin (vedoucí práce)
The purpose of this work is to create a model using spherical convolutional neural networks that can estimate the orientation of a camera from two inputs, where the first input is a panorama and the second input is a photograph capturing a specific part of the panorama. In other words, the task is to find where in the panorama, which is the first input, is located the photo, which is the second input. In addition to three created models that address this problem, six new datasets have also been created, which expand the currently available number of datasets whose photos are in equirectangular or stereographic format.
|
|
|
Camera Pose Estimation from Lines using Direct Linear Transformation
Přibyl, Bronislav ; Pajdla, Tomáš (oponent) ; Koch, Reinhard (oponent) ; Zemčík, Pavel (vedoucí práce)
This thesis is concerned with camera pose estimation from correspondences of 3D/2D lines, i.e. with the Perspective- n -Line (PnL) problem. Attention is focused on large line sets which can be efficiently solved by methods using linear formulation of PnL. Up to date, methods working only with point-line correspondences were known. Motivated by this, two novel methods based on the Direct Linear Transformation (DLT) algorithm are proposed: DLT-Plücker-Lines working with line-line correspondences and DLT-Combined-Lines working with both point-line and line-line correspondences. In the latter case, the redundant information reduces the minimum of required line correspondences to 5 and improves accuracy of the method. The methods were extensively evaluated and compared to several state-of-the-art PnL methods in various conditions including simulated and real-world data. DLT-Combined-Lines achieves results similar to or better than state-of-the-art, while it is still highly efficient. In addition, the thesis introduces a unifying framework for DLT-based pose estimation methods, within which the proposed methods are presented.
|
|
|
Real-time camera pose estimation for augmented reality
Szentandrási, István ; Slavík,, Pavel (oponent) ; Liarokapis,, Fotis (oponent) ; Herout, Adam (vedoucí práce)
Fiduciary markers form the base for camera pose estimation for many Augmented Reality applications, when a fast and robust solution is required. In this thesis an efficient marker-based camera pose estimation is outlined for Uniform Marker Fields with several real-world applications. The proposed algorithm is highly efficient and works in real time even on multiple platforms, including mid-range smartphones. On-screen markers are used to establish relative pose between devices for task-migration and information reaccess. In movie production, the described camera pose estimation as part of the chromakeying process, provides content creators real-time preview. Results show that the described detection algorithm performs comparably to other marker-based methods and it is several times faster. The implemented applications provide viable - cheaper and faster - alternative to existing solutions.
|
|
|
Camera Pose Estimation from Lines using Direct Linear Transformation
Přibyl, Bronislav ; Pajdla, Tomáš (oponent) ; Koch, Reinhard (oponent) ; Zemčík, Pavel (vedoucí práce)
This thesis is concerned with camera pose estimation from correspondences of 3D/2D lines, i.e. with the Perspective- n -Line (PnL) problem. Attention is focused on large line sets which can be efficiently solved by methods using linear formulation of PnL. Up to date, methods working only with point-line correspondences were known. Motivated by this, two novel methods based on the Direct Linear Transformation (DLT) algorithm are proposed: DLT-Plücker-Lines working with line-line correspondences and DLT-Combined-Lines working with both point-line and line-line correspondences. In the latter case, the redundant information reduces the minimum of required line correspondences to 5 and improves accuracy of the method. The methods were extensively evaluated and compared to several state-of-the-art PnL methods in various conditions including simulated and real-world data. DLT-Combined-Lines achieves results similar to or better than state-of-the-art, while it is still highly efficient. In addition, the thesis introduces a unifying framework for DLT-based pose estimation methods, within which the proposed methods are presented.
|
|
|
Real-time camera pose estimation for augmented reality
Szentandrási, István ; Slavík,, Pavel (oponent) ; Liarokapis,, Fotis (oponent) ; Herout, Adam (vedoucí práce)
Fiduciary markers form the base for camera pose estimation for many Augmented Reality applications, when a fast and robust solution is required. In this thesis an efficient marker-based camera pose estimation is outlined for Uniform Marker Fields with several real-world applications. The proposed algorithm is highly efficient and works in real time even on multiple platforms, including mid-range smartphones. On-screen markers are used to establish relative pose between devices for task-migration and information reaccess. In movie production, the described camera pose estimation as part of the chromakeying process, provides content creators real-time preview. Results show that the described detection algorithm performs comparably to other marker-based methods and it is several times faster. The implemented applications provide viable - cheaper and faster - alternative to existing solutions.
|
host ::
RSS.