Page 142 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

Section 4.7. A Comparative Study 119

of a higher computational complexity, the performance of the PRA can be
improved by increasing the number of iterations. This is also true when ex-
amining more peaks for the PCA.
Figure 4.13 compares the prediction quality of the four algorithms when
applied to the three test sequences AKIYO,FOREMAN, and TABLE TENNIS,at
di erent frame skips (and, consequently, di erent frame rates).
As expected, the DFA performs well for sequences with a low amount
of movement (AKIYO) and at low frame skips (i.e., high frame rates). For
sequences with a higher amount of movement (FOREMAN and TABLE TENNIS)
and also at high frame skips, the motion vectors become longer, the quality
of the Taylor series approximation becomes poor, and the performance of the
DFA deteriorates.
Due to its dense motion eld, the PRA has a superior performance for
AKIYO and a very competitive performance for FOREMAN and TABLE TENNIS.
The relative drop in performance for high-motion sequences and at high frame
skips may be due to a number of reasons. With longer motion vectors, there
is more possibility that the algorithm will be trapped in a local minimum
before reaching the global minimum. Also, the maximum number of iter-
ations may not be su!cient to reach the global minimum. However, in-
creasing the number of iterations will increase the complexity of the
algorithm.
In general, the performance of the PCA is somewhere in between that of the
DFA and PRA. The poor performance for AKIYO may be due to the spurious
peaks produced by the boundary and spectral leakage e ects. Such e ects may
be reduced by applying a weighting function to smooth the phase correlation
surface.
The best overall performance is provided by the BMA. It performs well
regardless of the sequence type and the frame skip. In fact, for sequences
with a high amount of movement (FOREMAN and TABLE TENNIS), the BMA
shows superior performance.
It is interesting at this point to concentrate on the PRA and BMA, for
two reasons. First, they achieved the best prediction quality performance in
the comparison. Second, they represent two di erent approaches to motion
estimation (pel-based and block-based, respectively). Figure 4.14 compares
the performance of the PRA and the BMA for the rst 50 frames of the
FOREMAN sequence at 25 frames=s. Two versions of the PRA are considered:
PRA, which is the same algorithm described earlier, and PRA-C, which is an
algorithm in which the update term is based on the causal part of an area of
5 × 5 pels centered around the current pel. Since PRA-C is based on causal
data, no motion overhead needs to be transmitted for this method. Due to the
high amount of motion in FOREMAN, the maximum number of iterations for
both pel-recursive algorithms was increased to 10.