|M.Sc Student||Alexandr Kaplan|
|Subject||Finding Epipolar Geometry from Two Color Images|
|Department||Department of Computer Science||Supervisors||Full Professor Shimshoni Ilan|
|Full Professor Rivlin Ehud|
In this thesis we present a novel method for finding the epipolar geometry from two images by using color information. This method consists of algorithms for feature detection, feature matching and epipolar geometry finding from a list of corresponding pairs. Our algorithm for feature detection is based on an existing gray-scale corner detector and extends it to color. The feature matching algorithm first finds the most significant colors in the area around the feature. After that, we compare the lists of image features in all the possible rotations of the images and look for a sharp peak in the function, which describes the criteria for correct matching. The epipolar geometry finding algorithm is based on RANSAC and the eight point algorithm. It minimizes a cost function, which computes the distance from points to their epipolar lines.
Our method is able to find the epipolar geometry from two color images with unrestricted
changes robustly. The tests we ran show that our algorithm can cope with very complicated cases, such as big rotations and large changes in viewing direction between the two images.
We confirm the correctness of our algorithm by providing the results of the running of the algorithm on many pairs of indoor and outdoor images.