The set of correspondences is produce by an automatic wide-baseline stereo matching tool, it contains both correct correspondences and mismatches. These must be separated, according to consistency with epipolar constraint:
- for each corresponding pair $i$ of points, compute symmetric epipolar distance
$d_i = 0.5\sqrt{d_{1,i}^2 + d_{2,i}^2}$
where $d_{1,i}$ is euclidean (!) distance of a point $u_{1,i}$ to the corresponding epipolar line $l_{1,i} = F^T u_{2,i}$ in the first image, and $d_{2,i}$ is euclidean distance of a point $u_{2,i}$ to the corresponding epipolar line $l_{2,i} = F u_{1,i}$ in the second image.
- choose appropriate threshold, e.g. $\theta=1$ px and separate correspondences to inliers ($d_i \le \theta$) and outliers ($d_i \gt \theta$).
- Show the inliers (red) and outliers (black) as a needle map over one image. (A 'needle' is line joining coordinates of corresponing point in both image, but drawn to a single image.) Consistent motion should be visible in the inliers.
- Select 20 (eg. random) correspondences from inliers, and draw to both images the points and the corresponding epipolar lines, in different but corresponding colours.