Graphics Reference
In-Depth Information
Robust Matching
Structure/Motion Estimation
Database Update
Fig. 2.2
View creation process
View creation proceeds in three steps (see Fig.
2.2
):
1. Robust matching: Correspondences are established between features in two or
more temporally local images, while enforcing geometric constraints.
2. Structure estimation: From the feature correspondences, three-dimensional point
structure is estimated and stored. The (optional) differential motion estimate is
used to determine the common scale of the structure.
3. Database management: The appearance model of the view (comprising a set of
feature descriptors) is added to a global database for later recognition.
2.5.1 Robust Matching
The interframe matching procedure for view creation first establishes putative corre-
spondences then partitions these correspondences into inliers (correct matches) and
outliers (incorrect matches) using geometric constraints.
Putative correspondences can be generated using only the feature descriptors, or
by taking advantage of any differential motion estimates supplied by other sensors,
such as wheel odometry. In the first case, each feature in the current image is paired
with the feature in a recent older image according to distance in the feature descriptor
space using a brute-force or approximate nearest neighbors (ANN) method. In the
second case, a motion estimate constrains the search for putative correspondences.
The nearest feature in descriptor space that also satisfies the corresponding epipolar
constraint is taken as a putative correspondence to the older feature.
Given a set of putative correspondences, geometric constraints are applied iter-
atively to eliminate outliers. If no prior on camera motion is provided, a starting
point for the procedure can be computed using RANSAC [
6
] and the five-point
