Graphics Reference
In-Depth Information
algorithm [
15
]. This yields an estimate of camera motion (up to scale) and a set of
inlier correspondences. When a prior differential motion estimate is available, it is a
sufficient starting point for iteration.
The iteration proceeds as follows:
1. An error threshold factor
r
is chosen as a multiple of the desired final acceptance
˜
threshold
r
.
2. Inliers are selected by finding all putative correspondences whose matches fall
within a threshold distance of the epipolar line, whose descriptor distances are
low, and whose depth estimates are positive. The epipolar threshold distance
for a feature with scale
s
is given by
r
˜
·
s
, modeling larger location uncertainty
associated with larger-scale features.
3. The motion estimate is refined by nonlinear maximum-likelihood estimation over
the current set of inlier correspondences.
4. The threshold factor
r
is decreased multiplicatively, and the process is repeated
˜
from step 2 until
r
˜
≈
r
.
We use this approximation to a standardM-estimator scheme (e.g., iterated reweighted
least squares with Tukey weighting) in order to reduce the computational cost on
embedded platforms.
2.5.2 Structure Estimation
Given feature correspondences between two views, bundle adjustment [
19
] is per-
formed over the reprojection objective function to yield joint estimates on structure
and camera motion. The scale is left unconstrained by the feature correspondences,
so the gauge freedom is eliminated by fixing the camera translation to unit mag-
nitude while performing the optimization. The scale is assigned to the view using
the differential odometry between the two views used for estimation. Further views
can be added to the optimization either at the point of view creation or upon later
observation. In this case of upgrading the structure, the camera translation magnitude
is constrained only between the first two views, and all six degrees of freedom vary
among the others. The previously computed parameter values are used as a starting
point in the new, larger optimization.
2.5.3 Database Management
A global appearance database is maintained to aid view recognition. When a new
view is created, its appearance model is added to this database.
The global database could take one of many forms, depending on the desired
appearance model representation. We describe a simple but effective approach here.
