Compression - 3DTV: Processing and Transmission of 3D Video Signals - page 78

Figure 3.6 Generic prediction structure for the MVC method
dynamically within the same frame. However, disparity among views is
more dependent on the distance between the cameras and can have much
greater values than the motion vectors. Furthermore, there generally exists
a global disparity between the views, which does not change in time, unless
the multiple cameras are moving.
Experiments have been performed with different multi-view sequences to
analyse the efficiency of interview prediction structures in [20]. It has been
shown that temporal prediction is the most efficient prediction mode for
all analysed sequences on average. The relationship between the temporal
and the inter-view prediction strongly depends on the scene complexity
and the temporal/spatial density. However, in the results shown in [20],
it is seen that inter-view prediction attains a quality improvement of
0.5 - 2 dB for most of the sequences over coding each view separately with
the hierarchical B-frame prediction. However, this gain changes from one
prediction structure to another. Furthermore, it is seen that when there is too
much disparity between neighbouring camera views, the encoder cannot
fully exploit the redundancies.
The basic prediction structures are shown in Figure 3.7. In Figure 3.7
(a), the simulcast coding structure is shown. This shows that no inter-view
prediction is used. In Figure 3.7 (b), inter-view prediction is used only for key
frames. In Figure 3.7 (c), inter-view prediction is used for non-key frames as
well as the key frames. The prediction structure denoted in Figure 3.7 (c) is
found to be the best performer (1.7 dB gain on average). However, it achieves
a small gain over the structure denoted in Figure 3.7 (b) for a much increased
complexity. It is interesting to note that the prediction structure with bi-
predictive coding applied to key frames performs worse in general when
compared to the scheme with no bi-predictive coding applied to key frames.
This is due to the QP cascading scheme used in hierarchical B prediction.
With QP cascading, there is less fidelity for lower hierarchy levels. For
instance, B Frames are encoded with a higher QP than P frames and I frames.