Page 174 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

Section 6.3. Long-Term Memory Motion-Compensated Prediction 151

where D INTRA is the SSD between the current macroblock and its INTRA
encoded reconstruction and R INTRA is the number of bits used to INTRA
encode the current macroblock. Similar de)nitions also apply for D INTER
and R INTER , but they are calculated by INTERencoding the current mac-
roblock. In both equations, MODE is a Lagrange multiplier related to the
quantization parameter QP using

MODE =0:85 × QP : (6.8)
2
The mode with the minimum cost function is chosen as the mode of
the current macroblock. Note that, in this case, a macroblock needs to
be encoded twice before being able to decide its mode. This increases
the complexity of the encoder. A more detailed description of this rate-
constrained motion estimation and mode decision method can be found
in Ref. 143.

MR This is a multiple-reference encoder with no rate constraints. Thus, it
is the same as SR, but it uses long-term memory motion-compensated
prediction.

MR-RC This is a multiple-reference rate-constrained encoder. Thus, it is the
same as SR-RC, but it uses long-term memory motion-compensated pre-
diction.
The preceding encoders were tested using the three QSIF test sequences
AKIYO,FOREMAN, and TABLE TENNIS. The frame skip parameter was set to 3
to achieve low bit rates. To generate a range of bit rates, the quantization
parameter QP was varied over the range 5–30 in steps of 5. This means that
each encoder was used to encode a given sequence six times. Each time, QP
was held constant over the whole sequence (i.e., no rate control was used).
The )rst frame was always INTRA encoded. The INTRA bits of the )rst frame
were included in the bit-rate calculations, and no header bits were generated.
All quoted results refer to the luma components of sequences. For MRand
MR-RC, sliding-window control was used to maintain a long-term memory of
size M = 50 frames. The VLC codewords in Table 6.1 were used to encode 1
the temporal components d t of the long-term motion vectors.
Figures 6.6, 6.7, and 6.8 show the R-D performance of the preceding
encoders for the three test sequences. Note that both single-reference and

1 For example, since d t = 4 is in the range (3:6), then according to Table 6.1 it will be encoded
using a 5-bit codeword. This codeword is derived as follows. With reference to the start of its
range, d t = 4 is represented by d t − 3=4 − 3 = 1. Thus, x 1 x 0 = 01 and the codeword is given by
0x 1 1x 0 0 = 00110.