Page 144 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

Section 4.7. A Comparative Study 121

Foreman @ 25 f.p.s. 5 Foreman @ 25 f.p.s.
42 10
40
38
4
10
36
PSNR Y (dB) 34 Motion bits
32 3
10
30
28 BMA
PRA BMA
PRAC 2 PRA
26 10
1 5 10 15 20 25 30 35 40 45 50 1 5 10 15 20 25 30 35 40 45 50
Frame Frame
(a) Reconstruction quality (b) Motion bits
Foreman @ 25 f.p.s. Foreman @ 25 f.p.s.
4 5
10 10
3
10 10 4
DFD bits Total bits
10 2 10 3

BMA BMA
PRA PRA
PRAC PRAC
1 2
10 10
1 5 10 15 20 25 30 35 40 45 50 1 5 10 15 20 25 30 35 40 45 50
Frame Frame
(c) DFD bits (d) Total bits = motion + DFD
Figure 4.14: Comparison between BMA and PRA motion estimation algorithms
The aim of motion estimation for video coding is to simultaneously min-
imize the bit rate corresponding both to the motion parameters (motion bits)
and to the prediction error signal (DFD bits). As illustrated in Figure 4.14, the
three algorithms represent three di erent tradeo s between prediction quality
and motion overhead. Due to its dense motion eld, the PRA has the best
prediction quality and, consequently, the least DFD bits. This is, however, at
the expense of a prohibitive motion overhead, which leads to a very high total
bit rate. The causal implementation of the PRA, PRA-C, clearly restricts the
method and signi cantly reduces its prediction quality. Thus, PRA-C removes
the motion overhead at the expense of an increase in DFD bits. In addition,
this causal implementation increases the complexity of the decoder. The best
tradeo is achieved by the BMA. It uses a block-based approach to reduce
the motion overhead while still maintaining a very good prediction quality.
This explains the popularity of this approach and its inclusion in video coding
standards.