engleski

Performance evaluation of the five-point relative pose with emphasis on planar scenes

We consider performance evaluation of the state-of the-art solution for recovering the relative pose between two calibrated views. Our focus is on planar scenes which are not tractable by algorithms which do not enforce the so-called calibrated constraint. The capability to cope with planar scenes has therefore been stressed as an important advantage of the novel five-point algorithm. However, we show that for planar and nearly planar scenes there is a considerable degradation of the five-point algorithm performance under noise. This is especially the case for sidewise motion, for which substantially better motion hypotheses can be obtained by homography decomposition. The differences are even greater when more than five points are available, since the accuracy of the homography approach scales better. We also note that, contrary to the previous claims, the five point algorithm is not a method of choice even in non-planar overconstrained contexts, since the performance of the classical 8pt algorithm can be greatly improved by equilibration. Thus, the results imply that the five-point algorithm is the best option only for non-planar scenes in minimal cases (as a hypothesis generator in a RANSAC scheme). At the price of a perhaps acceptable performance deterioration, the five point algorithm could be used for planar scenes as well, but only for prevalently forward motion.

relative pose; performance evaluation

nije evidentirano