• If x and x' are corresponding image points in two views, then
x' T Fx = 0. (2)
• 1' = Fx is the epipolar line corresponding to x (since x' T l' = 0). 1 = F T x' is
the epipolar line corresponding to x'.
• The epipoles satisfy the equations Fe = 0, F T e' = 0.
• F has 7 DOF — there are 9 matrix elements but only their ratio is significant,
which leaves 8 DOF. In addition the elements satisfy the constraint detF = 0
which leaves 7 DOF.
• Given the image correspondences of 7 points in general position, there are one
or three real solutions for F. Given the image correspondences of 8 or more
points in general position F is determined uniquely up to scale from the eight
linear equations provided by (2).
2.1.2 Computation of F for an uncalibrated image sequence
We assume here that the scene viewed by the moving camera is static. In an image
sequence (acquired at 25Hz) the constraints available are that features will generally
not have moved “too far” between frames, and will not have changed “too much”
in appearance. Algorithms have been developed which use these constraints and
no others in order to compute the epipolar geometry and match features between
two views — i.e. there is no a priori knowledge of motion or internal calibration.
Generally, the algorithm proceeds by extracting point features from each image
using interest operators (often termed “corner detectors”). These point features
are then matched automatically between views. The two constraints are utilised
in the matching by restricting correspondence search to a window centred on the
current position of the feature, and assessing matching strength by measuring the
cross correlation of intensity around the putative corner match.
Generally far more correspondences are obtained than the minimum of eight
required for a unique solution (typically there may be 100+ correspondences in a
512 x 512 image), and a least-squares type scheme is employed to improve the estima
tion of F. Mismatches severely disrupt such least squares schemes (and mismatches
will inevitably occur). Consequently there has been increasing use of robust tech
niques, particularly the RANSAC algorithm, to reject outliers (mismatches) during
the estimation.
The random sample consensus paradigm (RANSAC) is the opposite to conven
tional least squares techniques where as much data as possible is used to obtain a
solution. Instead, as small a subset as is feasible is used to estimate the parameters,
and the support for this solution measured. For example, in the computation of the
fundamental matrix, putative fundamental matrices are estimated using random
samples of seven correspondences. The distance of all points from their correspond
ing epipolar line (defined by the fundamental matrix) is then calculated, and if it is
Zisserman - 4