Determining the relative pose (rotation and translation) between two cameras or views of a scene is a fundamental problem in computer vision. It is common to compute this pose from matching points between two images, and it is well known that five such point correspondences are required. If eight correspondences are available, then a simple linear algorithm can be used.
However, these correspondences are not typically between simple points. In particular, there is often orientation information associated with each point. In this seminar I will show that this information can be used to create a novel four-point algorithm for relative pose problem, as well a linear seven-point solution. This reduction in the number of correspondences required generally leads to significantly faster pose estimation, with no loss of accuracy. Some specific cases are also identified where the new methods do not give an advantage (but are no worse than the original algorithms).
Allergy information: Contains geometry, linear algebra, quaternions, and traces of matrices. May contain nuts. Suitable for vegetarians.
Last modified: Tuesday, 10-Apr-2018 13:48:03 NZST
This page is maintained by the seminar list administrator.