|Title:||Closed-Form Solutions to Computation of Multiple-View Homographies|
|Abstract:|| Several fields of applications in computer vision require 2d-homography estimation based on point-correspondences between multiple images, which put into relation the 2d-projective transformations of each image in regard to a global reference coordinate frame. Due to noise in the measured point coordinates and eventual false matches on the point correspondences, it is impossible to find an exact solution for these homographies. Therefore - under certain assumptions about the type of noise - the aim is to find best approximates, notably Maximum Likelihood Estimators (MLE) for those homographies. Unfortunately the task of determining those MLE involves, amongst other processing steps, the optimization of a non-linear cost function which requires initialization. The quality or accuracy of the initialization also affects the quality and accuracy of the final result of the optimization. Hence, improving the initialization technique is of significant interest when it comes to high-precision mosaicing.|
So far, common initialization techniques either apply threading type methods to parallax-free scenarios or they apply batch type methods to sets of images with altering camera centers, although the concept of batch techniques allows the deviation of closed-form solutions. These closed-form solutions mainly differ from the threading type methods by the fact that they discard a whole level of detail from statistical point of view in order to linearize solving for an initial guess instead of heuristically withdrawing known and interpolating missing information.
This thesis introduces some of those closed-form solutions and investigates their behaviour in regard to a simple threading type solution with synthetic experiments and proves practical feasibility with a real example.
|Supervisor:||Adrien Bartoli, Pierre Fite-Georgel|
|Area:||Registration / Visualization, Computer Vision|