In recent years, the availability of texture mapping graphics hardware has increased, so that most desktop computers are able to display exceedingly realistic three-dimensional scenes. To achieve these new levels of realism it is important that textured models are obtained from photographic images.
Established geometric techniques exist for computing the motion between two or three views. The fundamental matrix describes the projective geometry between two views and the trifocal tensor describes the geometry of three views. In the past, sequences have been computed by linking together many fundamental matrices or tensors into long chains.
The trifocal tensor Hartley95 expresses the geometry of three views and can be computed from six points in three views, up to one of three projective ambiguities, as was shown by Quan Quan94 . Recently research into higher order tensors by Heyden Heyden98 has shown that the quadrifocal tensor (four views) is the highest order possible. For longer sequences the only improvement is the ability to average over a larger number of images.
Since closely spaced cameras have a small motion, it is important to use widely spaced views to improve the accuracy of the results. In this paper the trifocal tensor will be used to develop a method for computing the projection matrices of a sequence of images, so that all the projection matrices are in the same projective framework. This will be one of a family of solutions that all describe the projective geometry. The self-calibration technique by Pollefeys et al. Pollefeys98 is used to select one metric solution from the family of projective solutions.