next up previous contents
Next: Relating to other views Up: Updating the structure and Previous: Updating the structure and   Contents

projective pose estimation

For every additional view the pose towards the pre-existing reconstruction is determined, then the reconstruction is updated. This is illustrated in Figure 5.1.

Figure 5.1: Image matches ( ${\tt m}_{i-1},{\tt m}_i$) are found as described before. Since the image points, ${\tt m}_{i-1}$, relate to object points, ${\tt M}_i$, the pose for view $i$ can be computed from the inferred matches ( ${\tt M}, {\tt m}_i$). A point is accepted as an inlier if its line of sight projects sufficiently close to all corresponding points.
\begin{figure}%%[h]
\centerline{\psfig{figure=images/newaddview2.eps, width=15cm}}\end{figure}
The first step consists of finding the epipolar geometry as described in Section 4.2. Then the matches which correspond to already reconstructed points are used to infer correspondences between 2D and 3D. Based on these the projection matrix ${\bf P}_k$ is computed using a robust procedure similar to the one laid out in Table 4.3. In this case a minimal sample of 6 matches is needed to compute ${\bf P}_k$. A point is considered an inlier if it is possible to reconstruct a 3D point for which the maximal reprojection error for all views (including the new view) is below a preset threshold. Once ${\bf P}_k$ has been determined the projection of already reconstructed points can be predicted. This allows to find some additional matches to refine the estimation of ${\bf P}_k$. This means that the search space is gradually reduced from the full image to the epipolar line to the predicted projection of the point.


next up previous contents
Next: Relating to other views Up: Updating the structure and Previous: Updating the structure and   Contents
Marc Pollefeys 2002-11-22