Using the coupled self-calibration algorithm described in Section 6.3.1 it is possible to recover the metric structure of the different subsequences.
Once the metric structure of the subsequences has been recovered, the pose of the camera can also be determined for the viewpoints observing only planar points. Since the intrinsics have been computed, a standard pose estimation algorithm can be used. We use Grunert's algorithm as described in . To deal with outliers a robust approach was implemented .
Finally, it becomes possible to align the structure and motion recovered for the separate subsequences based on common points. Note that these points are all located in a plane and therefore some precautions have to be taken to obtain results using linear equations. However, since 3 points form a basis in a metric 3D space, additional points out of the plane can easily be generated (i.e. using the vector product) and used to compute the relative transform using linear equations. Here again a robust approach is used.
Now that all structure and motion parameters have been estimated for the whole sequence. A final bundle adjustment is carried out to obtain a globally optimal solution.