The procedure proposed in th eprevious section only relates the image to the previous image. In fact it is implicitly assumed that once a point gets out of sight, it will not come back. Although this is true for many sequences, this assumptions does not always hold. Assume that a specific 3D point got out of sight, but that it is visible again in the last two views. In this case a new 3D point will be instantiated. This will not immediately cause problems, but since these two 3D points are unrelated for the system, nothing enforces their position to correspond. For longer sequences where the camera is moved back and forth over the scene, this can lead to poor results due to accumulated errors. The problem is illustrated in Figure 5.2
The solution that we propose is to match all the views that are close with the actual view (as described in Section 4.2). For every close view a set of potential 2D-3D correspondences is obtained. These sets are merged and the camera projection matrix is estimated using the same robust procedure as described above, but on the merged set of 2D-3D correspondences.
Close views are determined as follows. First a planar-homography that explains best the image-motion of feature points between the actual and the previous view is determined (using Equation 5.2). Then, the median residual for the transfer of these features to other views using homographies corresponding to the same plane are computed (see Equation 5.1). Since the direction of the camera motion is given through the epipoles, it is possible to limit the selection to the closest views in each direction. In this case it is better to take orientation into account [47,76] to differentiate between opposite directions.