Page 224 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

Section 8.6. Simplex Minimization for Multiple-Reference Motion Estimation 201

(a) Original frame (b) Compensated using SR-FS (28.24 dB and
77,439 locations)

(c) Compensated using MR-FS with M =50 (d) Compensated using MR-3DSM with M =50
(31.31 dB and 3,871,950 locations) (31.04 dB and 72,532 locations)

Figure 8.12: Subjective quality of the motion-compensated 158 th frame of QSIF FOREMAN at
25 frames=s

of single-reference full-search SR-FS, and yet they still maintain the improved
prediction gain of multiple-reference motion estimation. This is also illustrated
in Figure 8.12, which shows the subjective quality of the motion-compensated
th
158 frame of FOREMAN. The uncovered background at the bottom-right corner
of the frame is poorly compensated using the single-reference algorithm SR-
FS (Figure 8.12(b)). This uncovered background is compensated with higher
quality using the multiple-reference algorithms (Figures 8.12(c) and 8.12(d)).
While the MR-FS algorithm achieves this improved prediction quality at the
expense of about 50 times increase in computational complexity, the MR-
3DSMalgorithm provides a similar improvement at no increase in computa-
tional complexity.