Translational registration proceeds in several stages. First, the
direction of translation (baseline) between each adjacent pair of
cameras is determined. Next, these pair-wise direction estimates are
assembled into a global set of constraints to find the relative
positions of all cameras. Finally, the relative positions are aligned to
absolute (earth-relative) positions by an optimal Euclidean
transformation.

Point features are extracted from each image, and
every possible correspondence (match) between points is evaluated.
Correspondences are ruled out according to hard geometric constraints,
such as positive depth and consistency of constituent edge directions.
What remains is a much smaller set of putative matches between features,
which is then used to refine the initial translation direction obtained
from the instrumentation.

When rotations are known, the epipolar geometry for a given camera pair
is simplified considerably. In particular, epipolar lines become arcs of
great circles on the sphere, and the epipole, or intersection of all
arcs, is the motion direction. This allows us to find the most likely
direction of motion by accumulating intersections between all possible
epipolar lines in a Hough transform space and finding the peak, or point
of maximum incidence.

The HT produces a fairly accurate (within a half degree or so) estimate
of the translation direction. A subsequent step refines this direction
by sampling from the distribution of all possible correspondence sets,
thus producing probabilistic correspondence. A Monte Carlo expectation
maximization (MCEM) algorithm is utilized which alternates between
estimating this correspondence distribution (E step), and finding the
direction given the current distribution estimate (M step).

Note that for a given pair of adjacent cameras, the translation can be
determined only up to scale; that is, we can only find the _direction_
of the motion, not its magnitude. Once this direction is found for each
camera pair, its [relative] magnitude can be determined by optimizing
over a set of linear constraints - namely, the set of all inter-camera
motion directions. This is achieved using constrained linear least
squares, the result of which is a set of mutually consistent camera
positions relative to an arbitrary frame of reference.

The final step is to register the resulting camera pose with respect to
an earth-relative coordinate frame. The entire set of cameras is
translated, rotated, and scaled according to the global transformation
that best aligns the estimated positions with the originally acquired
positions.