A SYSTEM FOR VIDEO OBJECT SEGMENTATION - Video Object Extraction and Representation: Theory and Applications - page 101

As shown in Figure 4.21, this sequence is problematic because of diffi-
culty of motion estimation. Most of the moving regions are homogeneous
and confuse the motion estimation (an aperture problem). Thus, most
information comes from the visual edges. Also, the white flagpole oc-
cludes the container ship and splits the initial object estimates into two
objects. In the current system, this type of error is unrecoverable. In the
future, high-level scene understanding may avoid these problems. Ob-
jective results (defined in Section 4.23) for the container sequence are
shown in Figure 4.24.

RESULTS FROM HALL-MONITOR SEQUENCE
The hall monitor sequence (frames 60-69) has one person walking
down the hall. The camera and background are stationary and the
person is walking away from the camera.
As shown in Figure 4.22, this sequence is the most troublesome for
our system. Since the figure is non-rigid, the bootstrap stage incorrectly
finds two differently moving regions for the single object and, in the cur-
rent system, we cannot recover from this problem. If we knew a priori
that the background was stationary, then we could rely upon change
detection more strongly. The adaptive or heuristic weighting of infor-
mation sources are in our future research. Most regions on the person
are homogeneous, giving little or no motion information. Unfortunately,
there are no ground truth sequences for the hall monitor sequence, so
no objective results could be calculated.

AN OBJECTIVE MEASURE OF
SEGMENTATION QUALITY
In recent years, a major difficulty in evaluating results of content-
based video processing is the lack of good objective measures. We in-

→

troduce a simple measure called the 2-D quality vector, Q , defined in

Eq. 4.31 below:

〉

〈 || P truth ||

∩ ∩

|| P truth

P result ||

|| P truth

P result ||

,

→

, [0, 1] × [0, 1] (4.31)

Q =

|| P result ||

where P truth are the pixels in the ground truth segmentation and P result
are the pixels in the segmentation result. The components of the quality
vector are called, respectively, the content and coverage percentage.
As shown in Figure 4.23,

→

Q allows us to easily interpret the results

from segmentation algorithms. An ideal segmentation algorithm would
give results in

→

Q of < 1,1 > and any refinement to the segmentation