Recovering the epipolar geometry.

  The algorithm proceeds in three steps:

1. Firstly, we find a set of candidate matches between the two images using the only available information, i.e. the feature vector (see equation 3) computed for each point to match on the normalized images (see section 2.3). The vectors are compared using the method described in section 4.1. Then we eliminate the matching ambiguities using the relaxation algorithm described in section 4.2.

2. From the set of matches now established, we estimate the fundamental matrix F using a robust linear method with the help the Last Median of Squares regression (LMedS) [16] [10].

3. Finally, we can detect the outliers in the candidate matches by exploiting the geometric constraint found in the previous step, i.e. the epipolar geometry. For a given point in the first image, we search for the points in the second image which are near the epipolar line F and for which is near the epipolar line .