@INPROCEEDINGS{Botterilletal2011c,
author = {Tom Botterill and Steven Mills and Richard Green},
title = {Refining essential matrix estimates from RANSAC},
booktitle = {In Proceedings of Image and Vision Computing New Zealand},
year = {2011},
abstract={To estimate the relative pose of two cameras from
outliercontaminated feature correspondences, the essential matrix
and inlier set is estimated using RANSAC, then this estimate
is refined to minimise an error function on the correspondences.
This paper evaluates several refinement methods which minimise
functions of Sampson’s error. All perform well on large sets of
correspondences or when inlier rates are high, but many perform
poorly or fail when the inlier set found by RANSAC is small; this
is shown to be because the inlier sets contain remaining outliers,
while missing some inliers.
The most accurate solutions are given by minimising the robust
BlakeZisserman function of Sampson’s error, although this
provides only a minimal improvement in accuracy compared with
least squares refinement. The most reliable results are given by
nonlinear optimisation constrained to the essential manifold. An
efficient parametrisation of the essential manifold as a quaternion
and a unit vector is described; applying Iteratively Reweighted
Least Squares combined with LevenbergMarquardt optimisation
on this manifold takes typically less than one millisecond..}
}
