When reconstructing 3D environments, i.e. rooms or objects, from range images and photographs it is in general impossible to acquire the entire information within one view. Consequently, one has to address the problem of ``where to look next'', which is of particular importance in automatic reconstruction. Previously, this problem had mainly addressed the reconstruction of small objects with a very limited number of possible parameter sets for the next view, thus the approaches cannot easily be extended to large and complex environments. When reasoning where to observe next during the reconstruction of built environments, two main prerequisites have to be fulfilled: 1) the already reconstructed part of the scene has to be efficiently represented and 2) different types of sensors have to be integrated.
UOE and the JRC are developing techniques that reason about where to observe next and how to allocate resources in order to minimize the operational costs of the reconstruction. JRC's technique is occlusion driven and selects the ``next best view'' by choosing a location that satisfies ``hard constraints'' ( e.g. collision avoidance, visibility of occlusions) and optimizes a purposely defined criterion, thereby taking the quality of the model into account. UOE's technique is based on direct optimization of quality of surface and reduction of unseen surface area.
At the left is an example of JRC's results showing the estimated best next capture position on the Bornholm church data, and the right shows some results from UOE's algorithm, showing convergence of scanned area as a function of view number (on a different scene).
Preliminary experiments using the two algorithms have been done and published, and JRC and UOE are now collaborating on a comparison of the algorithms.
J. M. Sanchiz, R. B. Fisher ``A next-best-view algorithm for 3D scene recovery with 5 degrees of freedom'', Proc. British Machine Vision Conference BMVC99, Nottingham, pp 163--172, September 1999.