Page 219 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

196 Chapter 8. The Simplex Minimization Search

Table 8.7: Computational complexity within MPEG-4 in terms of searched locations=MB. Repro-
duced from Ref. 175
Sequence Object FS SMS DS NSS
Bream VO1: Fish 1,017.2 16.7 21.4 33.0
Coast Guard VO0: Water 1,021.8 16.4 18.0 33.0
VO1: Small boat 1,004.7 16.8 17.5 32.9
VO2: Big boat 1,010.0 17.7 19.1 33.0
VO3: River bank 1,020.5 13.1 19.2 32.9
Container ship VO0: Water 1,023.4 12.7 13.1 33.0
VO1: Ship 1,023.7 11.4 13.5 33.0
VO2: Small boat 1,014.2 14.7 15.9 32.9
VO3: Land (fg) 1,024.0 9.4 13.0 33.0
VO4: Sky+Land (bg) 1,024.0 13.9 13.1 33.0
VO5: Flag 1,012.3 19.1 15.6 33.0
News VO0: Background 1,024.0 9.3 13.0 33.0
VO1: Dancers 1,024.0 15.4 16.3 33.0
VO2: News readers 1,022.9 9.8 13.1 33.0
VO3: Text 1,024.0 9.0 13.0 33.1
Stefan VO0: Stefan 1,002.4 21.8 22.2 32.9
Minimum 1,002.4 9.0 13.0 32.9
Maximum 1,024.0 21.8 22.2 33.1
Average 1,018.3 14.2 16.1 33.0

8.6 Simplex Minimization for Multiple-Reference
Motion Estimation

As already discussed, MR-MCP achieves signi:cant prediction gains, but at
the expense of a signi:cant increase in computational complexity. This is
illustrated in Figure 8.10 for the FOREMAN sequence at 8:33frames=s. This :gure
was generated using the same simulation conditions described in Section 6.3.2.
Figure 8.10(a) shows the prediction quality (in terms of PSNR Y in decibels)
as a function of multiframe memory size (in frames), whereas Figure 8.10(b)
shows the computational complexity (in terms of searched locations=frame).
It is clear that increasing the memory size M increases the prediction qual-
ity. This is, however, at the expense of a linear increase in computational
complexity. The aim of this section is to design fast long-term memory block-
matching algorithms that can reduce computational complexity but at the same
time maintain the prediction gain of multiple-reference motion estimation.