Digital Signal Processing Reference
In-Depth Information
6. RESULTS AND ANALYSIS
We analyze the advantages and disadvantages for the system by run-
ning the system on standard MPEG-4 sequences [Sikora, 1997]. We
compare our results to ground truth results through an objective seg-
mentation quality measure in section 4.23. In our work, quality of results
was paramount and we did not optimize for running time. Running time
of the system varies, but is on the order of 10 hours for 10 frames.
THE MPEG-4 TEST SEQUENCES
The three sequences that are used in our test are from the standard
MPEG-4 test sequences: coastguard, container and hall monitor. Each
of the sequences are raw CIF format, 320 by 200 pixel resolution frame at
30 frames per second. The sequences are in color with YUV color space.
Since the Y components approximate image intensity, we use only the Y
channel in our calculations to approximate the grayscale images. We run
our system on 10 frame subsequences of the video that conform to the
simplifications in Section 2.. The results from these test sequences are
shown as follows: Figure 4.18 are results from selected frames, Figures
4.20, 4.21 and 4.22 are complete results from the various test sequences,
objective results are shown in Figure 4.24.
RESULTS FROM COASTGUARD SEQUENCE
The coastguard sequence (frames 30-39) has two foreground objects:
the larger coastguard ship and a smaller motorboat with its wake. A
short description of sequence follows: a large coastguard boat enters the
screen from the left and moves right, a small motorboat moves to the
right with water and shore in the background. Within this subsequence,
the camera pans with the motorboat, while the coastguard boat moves
to the right.
As shown in Figure 4.20, this sequence has particular qualities that
are best suited for our algorithm. The foreground objects are rigid and
their movements are orthogonal to the camera view, assuring that the
object projection remains in the same size and shape except for its dis-
placement. Thus, the temporal smoothness criterion works well in this
case. Visually, the objects stand out from the background, allowing
the edge detection algorithms to pick up the boundaries. The homoge-
neous background has a mixed effect on our calculations: it lessens noise
on the edge detection, but the determination of background motion is
more difficult. Our bootstrap stage finds the three objects in the video
sequence. Since the optical flow is accurate, the bootstrap finds the ob-
jects in question, the large coastguard boat and the small motorboat,
Search WWH ::
Custom Search