Information Technology Reference
In-Depth Information
Furthermore, inferring the occlusion relationship can be fulfilled by comparing the
match scores
pe
of each point
p
of the occluding edges. This is because an
occluded region only shares the same motion layer with one of the profiles of an oc-
cluding edge. The smaller match score corresponds to the real occlusion region. This
implies that one side of an occluding edge with the smaller match score is behind the
other side, since it involves the occlusion region. In terms of the profile labeling of the
occluding edge points, we can therefore find ordinal depth.
(
),
e
(
p
)
1
2
The
Post-processing procedure
is summarized as follows:
(1)
Extracting the pending areas
W
on each frame;
(2)
Refining the corresponding points
p
′
on the next frame (
i
+1), and then re-
compute the match scores
epepep
;
(3)
Applying the polysegment algorithm to the match scores of
W
for detecting
the points of layer boundaries;
(4)
Merging the spurious regions for the continuous boundaries of the motion
layers;
(5)
Determining the occluding edges in terms of
M
O
;
(6)
Extending the profiles of the occluding edge points to the known motion
layers for the profile labeling;
(7)
Comparing the match scores
( , ( , ()
0
1
2
epep
of the occluding edge points
p
for
( , ()
1
2
depth ordering, i.e.
min{
epep
corresponds to the occluded region.
(
),
(
)}
1
2
This post-processing procedure and the subspace segmentation approach described
in section 2 constitute a complete algorithm of motion segmentation. Note that in
our algorithm, the estimation of all the motion models in a scene is undertaken at
the first procedure (i.e. subspace segmentation method), and the detection of the
layer boundaries and depth ordering are carried out at the second procedure (i.e.
post-processing procedure). This is different from the previous approaches. Usu-
ally the motion model estimation was mixed with the later processing. This makes
the algorithms complicated and the implementation difficult.
4 Experiments and Analysis
Our algorithm was tested on several image sequences. In this section, the three
results of the 'flower garden', 'Susie calling' and 'table tennis' are presented. All
programs have been implemented on the MatLab platform using a publicly avail-
able package—GPCA-PDA [16] and the optical flow code in [15]. All the image
sequences used in our experiments are available at [18].
Flower Garden
In this experiment, we applied our motion segmentation approach to the flower
garden sequence of resolution 175×120 pixels. The tree trunk in front of a garden
Search WWH ::
Custom Search