Graphics Reference
In-Depth Information
As quality measure, we used a weighted sum of the average PSNR values for the
luma and chroma component, defined by
PSNR
YUV
D
.6
PSNR
Y
C
PSNR
U
C
PSNR
V
/=8;
(3.3)
where PSNR
Y
, PSNR
U
, and PSNR
V
represent the average PSNR values, averaged
over the pictures, for the individual color component. For video with a bit depth of
8 bit per sample, the picture PSNR for a particular color component is given by
0
s.x; y/
s
0
.x; y/
2
1
W 1
X
H 1
X
1
255
2
W
H
@
A
;
PSNR
D
10
log
10
(3.4)
xD0
yD0
where s and s
0
represent the W
H sample arrays of the original and reconstructed
color component, respectively. By measuring the reconstruction quality using the
PSNR
YUV
and the average bit rate of the generated bitstreams for ten different
quantization settings, we obtained a rate-distortion curve for each configuration
and sequence, which characterizes the coding efficiency. In order to express the
difference in coding efficiency between two configurations as a single number, we
calculated an average bit-rate saving. Therefore, we interpolated the rate-distortion
curves of the two considered configurations for a test sequence in the logarithmic
domain using cubic splines with the “not-a-knot” condition at the border points and
determined the average bit-rate saving for the sequence by numerical integration
with 1,000 equal-sized subintervals.
6
Finally, the average bit-rate savings for the
application scenarios given in the following were obtained by averaging the bit-rate
savings over the test sequences.
In a first experiment, we investigated the impact of supporting different block
sizes for signaling motion parameters. In order to exclude the effect of different
transform sizes, the maximum TU size was set to 4
4 for all investigated
configurations. As reference configuration, we used a setting in which only 16
16
PUs are enabled, similarly as in H.262
j
MPEG-2 Video. Based on this reference
setting, we successively enabled additional PU sizes as summarized in Table
3.3
.It
can be seen that also for high-resolution video, supporting PU sizes smaller than
16
16, as has been done in developing H.263, MPEG-4 Visual, and H.264
j
MPEG-4 AVC, improves the coding efficiency. However, the biggest gain is
achieved by additionally supporting PU sizes larger than 16
16 luma samples.
While most of the coding gain relative to the reference configuration can already
be obtained by enabling all square PUs from 4
4 to 64
64 luma samples,
the additional support of rectangular PUs with symmetric and asymmetric CU
partitioning modes further improves the coding efficiency. In order to include design
6
The employed approach for calculating an average bit-rate saving between two rate-distortion
curves represents a generalization, that is also applicable to experimental data with more than four
rate-distortion points, of the often used Bjøntegaard Delta bit rate (BD-rate) [
3
].