Graphics Reference
In-Depth Information
that can be reliably located in other images of the same environment, and that feature
matches ideally result from the image projections of the same 3D scene point.
We can apply any of the detection and matching methods from Chapter
4
.If
the images are generated by video frames close together in time and space (which
is usually the case), we typically select single-scale Harris corners (since a square of
pixels is likely to remaina square inan image takena fractionof a second later) anduse
a fast matching algorithm like the KLT tracker described in Section
4.1.1.2
and [
442
].
If the images are taken further apart (which often occurs in hierarchical methods for
long sequences, discussed in Section
6.5.1.2
), then an invariant detector/descriptor
like SIFT will give more reliable matches (but will also be slower). On a closed set
or green-screen environment, artificial markers (e.g., gaffer-tape crosses or more
advancedmarkers as discussed in Section
4.5
) can be added to the scene to introduce
reliably trackable features. Figure
6.1
illustrates tracked features in these various
scenarios.
High-quality matchmoving relies on the assumption that matched image fea-
tures correspond to the same 3D scene point. Therefore, estimating the fundamental
matrix (Section
5.4.2
) and automatically removing matches inconsistent with the
underlying epipolar geometry is very important. Regardless of how the features are
obtained, they should generally be visually inspected and edited for correctness, since
Figure 6.1.
Tracking features for matchmoving. Top row: When the images are close together,
single-scale Harris corners can be easily detected and tracked. Middle row: When the images
are further apart, the SIFT detector/descriptor can be used for wider baseline matching. Bottom
row: In a green-screen environment, gaffer-tape crosses and artificial markers can be placed on
surfaces to aid feature detection and tracking.
