Page 170 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

Section 6.3. Long-Term Memory Motion-Compensated Prediction 147

QSIF Foreman QSIF Foreman
0.7 0.8
M=1, Skip=1 M=1, Skip=1
M=50, Skip=1 M=50, Skip=1
M=50, Skip=4 0.7 M=50, Skip=4
0.6
0.6
0.5
0.5
p(C dx −L dx ) p(C dy −L dy ) 0.4
0.4
0.3
0.3
0.2
0.2
0.1 0.1
0 0
−30 −20 −10 0 10 20 30 −30 −20 −10 0 10 20 30
C dx −L dx C dy −L dy
(a) Distribution of the difference between (b) Distribution of the difference between
the horizontal component, d x, of the current the vertical component, d y, of the current
vector and its left neighbor vector and its left neighbor
QSIF Foreman
0.6
M=50, Skip=1
M=50, Skip=4
0.5
0.4
p(C dt −L dt ) 0.3
0.2

0.1
0
−50 −40 −30 −20 −10 0 10 20 30 40 50
C dt −L dt
(c) Distribution of the difference between
the temporal component, d t, of the current
vector and its left neighbor
Figure 6.4: Highly correlated long-term memory block-motion )eld

6.3.2 Prediction Gain
This subsection evaluates the prediction gain achieved by LTM-MCP. All re-
sults were generated using full-pel full-search long-term memory block match-
ing with blocks of 16 × 16 pels, a maximum allowed displacement of ± 15
pels, SAD as the distortion measure, restricted motion vectors, and original ref-
erence frames. All quoted results refer to the luma components of sequences.
Figure 6.5 shows the performance of LTM-MCP when applied to the three
QSIF sequences AKIYO,FOREMAN, and TABLE TENNIS with di&erent memory
sizes and di&erent frame skips. It is immediately evident from this )gure
that signi)cant prediction gains are achieved when utilizing more than one