Randomising RANSAC in Wide-Baseline Stereo

J. Matas
CMP Prague, Czech Republic and CVSSP U. of Surrey, UK

We first introduce the notion of a singular and measurement regions. A singular region is a subset of an image that is a projection of a part of a scene having a distinguishing property that allows its detection over a range of viewing and illumination conditions. A measurement region is defined by a transformation invariant construction from a singular regions.

We construct two wide-baseline algorithms based on detection of singular regions. Correspondence and epipolar geometry is obtained from pairs of views taken from significantly different viewpoints.

In a second part of the talk we study how randomisation of the verification step of RANSAC can speed up the process of epipolar geometry computation. An optimal parameter of one class of statistical tests implementing the randomisation is found. Aproximate formulae for the speed-up factor obtained by the test are derived. .