i
i
i
i
i
i
i
i
9.3. Mosaicing and Making Panoramic Images 229
Of course, getting the iterative algorithm started is no easy matter. When
building a panorama, it is likely that image I
n
will have a lot of overlap with
I
n−1
, and so starting with the R that placed I
n−1
in the panorama is probably
the best way to generate an initial guess. However, if the overlap between two
successive frames is small (less than 50%), an alternative strategy of determin-
ing R, or at least getting an initial guess for it, must be followed. The two
most successful appr oaches are:
1. Hierarchical matching. Small subsampled and smoothed versions of I
n
and I
c
are made, e.g., images may be downsampled from 512 ×512 to
32 × 32 as a first guess. In these smaller images, the overlap appears
greater and, therefore, easier to determine R.OnceanR has been
determined, a larger subsampled image (e.g., 256 × 256) is used and
the least squares algorithm applied so as to refine the best estimate
for R.
2. Phase correlation [10]. This is most useful when the overlap is very
small. It is generally used to estimate the displacement between two
planar images. For example, two aerial photos can be montaged using
this approach. In it, a 2D Fourier transform of the pair of images is
made. The phase difference between them at each frequency is deter-
mined. An inverse transform is applied to the difference signal. Back
in the spatial domain, the location of a peak in the magnitude shows
where the images overlap.
In cases where I
n
has no overlap with I
c
, I
n
will have to be discarded
(temporarily) and reconsidered later after other images which may overlap
with it have been added into the mosaic. All in all, producing panoramic
mosaic images from video footage shot on a basic camcorder is no easy task.
Once panoramic images have been obtained, they can be used in a num-
ber of different ways. For example, McM i llan and Bishop [13] have demon-
strated how two panoramic images acquired from camera centers located
about 20 inches apart can generate parallax motion in the viewer.
9.3.1 QuickTime VR
Apple’s QuickTime movie file format was introduced in Chapter 5 in the
context of a container for storing compressed video and movies. Within a
QuickTime movie, there can be multiple tracks, with each one storing some