Page 206 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

Section 8.5. Simulation Results 183

highest, second highest, and lowest, respectively. The search is terminated if
the following condition is satis:ed:

(|h x − l x |≤ a) ∧ (|h y − l y |≤ a) ∧ (|s x − l x |≤ a) ∧ (|s y − l y |≤ a); (8.8)

where a is the search accuracy and ∧ is the logical AND operator. In other
words, the search is terminated if the two highest (in terms of BDM value)
vertices of the simplex become neighbors to the lowest vertex. This criterion
was derived from the way the SMmethod works. As shown in Figure 8.2,
when the method converges to a minimum, the contraction operation starts
pulling all the vertices toward the minimum vertex. The main advantage of
this criterion is that it does not depend on a threshold.

8.4.4 Motion Vector Re1nement

The main disadvantage of the preceding termination criterion is that it is not
based directly on the function to be minimized, i.e., the BDM. As a result,
the search may sometimes converge to a suboptimal point. Experimental re-
sults show that in most cases this suboptimal point is in the neighborhood
of the global minimum. An extra step is therefore added to the search in
which the motion vector produced by SMis re!ned by searching its eight
nearest neighbors. Note that this does not signi:cantly increase the com-
plexity of the search, because most of those neighbors have already been
searched.

8.5 Simulation Results

8.5.1 Results Within an Isolated Test Environment

In this set of simulations, motion is estimated and compensated using original
reference frames. In e+ect, this is equivalent to lossless DFD coding. This is
particularly important for a fair comparison between di+erent algorithms on a
frame-by-frame basis, since poor prediction of one frame does not propagate
to, and a+ect the prediction of, the next frame. Hereafter, the term isolated
test environment will be used to refer to this test condition.
All results in this subsection were generated using blocks of 16 × 16 pels,
a maximum allowed displacement of ±15 pels, SAD as the distortion mea-
sure, restricted motion vectors, and full-pel accuracy. Motion vectors were
coded predictively using the prediction method and the VLC table of the
H.263 standard. All quoted results refer to the luma components of
sequences.