A major issue with all reconstruction techniques is their reliance on high quality, essentially outlier-free input data. Robust estimation methods cannot be directly applied because the increased computation involved is prohibitive. The alternative that has recently been proposed is to select a set of feature matches that are globally consistent, in the sense of satisfying the rigidity constraint. Matching of point features between a stereopair of images while enforcing rigidity in general involves computing the fundamental matrix [ 19 , 10 ]. In the case of an affine camera and 2D shape, we can instead compute the 2D affine transformation between the two images. can be written as:
where
and
define the transformation. We follow [
19
] and apply the RANSAC algorithm of Fischler & Bolles [
6
] to compute a large subset of feature matches consistent with a single
set of 2D transformation parameters. Details of the matching algorithm
may be found in [
11
].
Temporal matching proceeds in the same way as stereo matching, except
that we match the previous
structure
estimates
to the new image, rather than the previous feature set. We thus employ a
matching algorithm which fits the affine camera parameters
directly, given the estimate of the structure
and the new image feature positions
.
Adrian F Clark
