Graphics Reference
In-Depth Information
Figure 6.2.
A false corner introduced by a coincidence of perspective between two surfaces:
one in the foreground and one in the background. From the positions of the chimney and the
corner of the background wall in the two images, we can see that the white feature points don't
correspond to the same 3D location, even though the pixel neighborhoods are almost identical.
a bad feature track can throw off the rest of the matchmoving process. Furthermore,
it's necessary to remove features that may appear mathematically acceptable (i.e.,
low tracking error and consistent with the epipolar geometry) but nonetheless can
cause problems for matchmoving. For example, Figure
6.2
illustrates an example of
a mathematically reasonable feature match introduced by a coincidence of perspec-
tive. A “corner” has been detected by the visual intersection of two different surfaces:
one on the foreground building and one on the background building. As the camera
moves, so does the apparent “corner,” but it does not correspond to a fixed 3D point.
These kinds of
false corners
need to be removed before matchmoving.
In addition, an implicit assumption about the input video in matchmoving is
that we're estimating the camera motion with respect to a stationary background.
Even though a feature detectormay find high-qualitymatches on foreground objects,
these should not be used for matchmoving if they undergo independent motion
compared to the larger environment. For examples, features on pedestrians and cars
moving down a street should be removed (even if they generate high-quality tracks),
leaving only features on the stationary road and buildings. Such situations can also
be detected using a robust estimate of the epipolar geometry, but somemanual effort
may be required to ensure that the right group of matches is used (for example, a large
moving foreground object might generate many feature tracks and be mistaken for
the “scene”). Finally, features only visible in a few frames (short tracks) are generally
not very useful for matchmoving and can be deleted.
While the previous paragraphs addressed the removal of bad or confusing features,
we can also
add
new features to the image sequence in several ways. First, the feature
detector should always be generating new matches — especially as new parts of the
scene come into view at the edges of the image as the camera moves. In natural,
