A SYSTEM FOR VIDEO OBJECT SEGMENTATION - Video Object Extraction and Representation: Theory and Applications - page 84

homogeneous movement represent a conservative core regions of objects
and bootstrap our surface optimization.
We use PCA to simplify our motion field [Kung et al., 1996] [Lin
et al., 1996]. As mentioned in Section 2.4, we introduced a simple con-
cept called motion paths that groups pixels by their correspondence of
projected motion. Assuming pixels from the same objects are moving
together, we can find the motion paths of different objects by the princi-
pal component of their projected motions. For example, for a 10 frame
sequence, we derive a 20 length vector made up of the 10 x-y compo-
nents of the projected motion for each motion path. After estimating the
correlation matrix from the vectors of all motion paths, PCA extracts
the first eigenvector from the correlation matrix of the dataset and the
magnitude of the first eigenvector in each motion path. Next, we spa-
tially cluster the PCA of our motion paths. Coupling our PCA with the
initial path positions, the motion paths project into 3-D feature vector
space. Since initial path positions are mutually exclusive, the projection
of the motion paths into this feature space form a functional surface in
3-D feature space. After smoothing the surface, the peaks and saddles
on this surface correspond to regions of homogeneous movement. Us-
ing estimated partial derivatives of this surface, we cluster these motion
paths by their peak / saddle regions and derive initial conservative surface
estimates.
BOOTSTRAP RESULTS
As shown in Figure 4.12, the bootstrap stage yields conservative esti-
mates for initialization of our video object segmentation system. When
our extraction of a motion field is accurate, the bootstrap stage finds
core regions of objects by their homogeneous motion well and robustly.
These regions become our conservative surface estimates, S initial , and
are suitable for surface initialization, multiple object separation and our
containment , locality, and Voronoi Ordered Space components of E object
However, since we use only motion in our object detection, we cannot
find objects that move with the background.
The results from the MPEG-4 test sequences show the advantages
and disadvantages of the bootstrap stage. For the coastguard sequence,
we see that our bootstrap does its job quite well: the three foreground
objects (large coastguard ship, smaller motorboat and the wake from the
motorboat) are found. In the container sequence, the bootstrap stage
detects the two objects (the small boat and the large container ship).
Unfortunately, due to the occlusion of the white pole, the container ship
is split into two distinct objects. In the current system, we cannot join
these two objects back together, showing the need for higher-level anal-