Information Technology Reference
In-Depth Information
Video sequences are encoded using the MPEG-4 MoMuSys codec, which sup-
ports binary shape coding. Four standard 3D test sequences 'Orbi', 'Interview',
'Ballet' and 'Break dance' have been used for experiments. The 'Orbi' and 'Inter-
view' sequences are of standard definition (720 x 576) resolution, while the 'Bal-
let' and 'Break dance' sequences are of high definition (1024 x 768) resolution.
For these last two multi-view sequences, the fourth camera view and the corre-
sponding depth map computed from stereo are utilized in this experiment [35].
Each sequence is segmented into foreground and background objects, using the
segmentation technique discussed in section 3B, and each object is encoded sepa-
rately. The prioritization module then estimates the expected distortion and the
packet allocation to individual bit streams is changed accordingly. In the WLAN
scenario, each stream is classified as a separate traffic flow from the access point
to the receiver. Experiments have been conducted for a range of channel condi-
tions and quality is compared with standard MPEG-4 object based video coding.
4.2 Performance Evaluation
The algorithm's performance is shown in terms of average frame PSNR vs
AvPLR in Fig. 5. 90 frames of the 'Orbi' and 'Interview' sequence, and 70 frames
of the 'Ballet' and 'Break dance' sequence have been used to make one test se-
quence of each type. Each point of the graph has been taken as the average of 30
runs of the decoder for each sequence. Each sequence consists of three sub bit
streams, one for the depth information and the other two for prioritized objects.
The streams are transmitted on different transmission flows. Av PLRs for each
flow in different test cases is shown in Table 2.
The performance gain for the 'Break dance' and 'Ballet' sequences is signifi-
cantly higher than other sequences, which is due to the nature of the video content.
The 'Interview sequence', for example, has very little movement in it and the gen-
eral perception of the sequence is static, which, therefore, leaves very little room
for optimization. This is demonstrated by fairly high PSNR values even for very
high packets error rates for the non prioritized streams as well. The 'Break dance'
sequence on the other hand has very quick movement in it and video quality is
highly vulnerable to transmission errors. A significant gain in performance is ob-
served with this sequence, which makes the scheme suitable for use with sensitive
data transmission.
The QP values used for QP optimization are 12, 16, and 20 for the background
object, and 12, 14, and 16 for the foreground object. The optimal value of QP for
each object is used to encode that particular frame. The QP values used for stan-
dard MPEG-4 that is used in comparison are 10 for I frames and 12 for P frames.
Fig. 6 shows selected frames from simulations with test case E. The selected
frames provide a visual indication of the differences seen in the objective results.
