Projection matrix decomposition in AR - a study with Access3D

Access3D is an AR application developed for industrial maintenance. An important feature of Access3D is that, in addition to visualization, it allows the user to inquire further information, directly from the AR view through the hyperlinks. In Access3D, we use VRML for visualization, the advantage is the direct database inquiry through the hyperlinks. On the other hand, VRML requires explicit camera parameters for the virtual camera modeling. Explicit camera calibration is often obtained from projection matrix computation (PMC) followed by projection matrix decomposition (PMD). The PMD is known to be numerically bistable, this is a problem of great importance and research interest in computer vision. Different methods, including data normalization have been introduced for finding numerically stable algorithms. Our analysis shows that even if data normalization is applied to the PMC step, a non-favorite distribution of the 3D points may still cause error in the results and numerical instability. This paper provides a look at the numerical instability of PMD from a different point of view.