Graphics Reference
In-Depth Information
Tabl e 5. 2
Average bit rate savings of HEVC merge and skip mode using
HM8.0
BD-rate [%]
Class
Sequence
RA-Main
LB-Main
LD-Main
A
Traffic
8:8
n/a
n/a
(2560 1600)
PeopleOnStreet
6:5
n/a
n/a
Nebuta
1:5
n/a
n/a
SteamLocomotive
11:1
n/a
n/a
B
Kimono
9:6
7:8
5:4
(1920
1080)
ParkScene
7:3
7:0
5:6
Cactus
11:7
8:8
6:9
Basketball Drive
9:1
8:1
6:1
BQTerrace
10:2
10:3
5:9
C
Basketball Drill
7:5
7:6
6:1
(832
480)
all
9:2
7:5
5:7
PartyScene
4:8
3:4
2:5
RaceHorses
3:9
4:3
3:3
D
Basketball Pass
6:1
5:1
4:0
(416
240)
BQSquare
6:9
3:3
2:5
BlowingBubbles
6:1
3:7
3:1
RaceHorses
4:4
3:9
3:6
E
FourPeople
n/a
11:4
8:5
(1280
720)
Johnny
n/a
20:0
16:4
KristenAndSara
n/a
15:0
11:2
Average
7.3
8.0
6.0
syntax. The software used for this experiment is version 8.0 of the HEVC test model
reference software (HM) [
13
] with the random access main, low delay B and P main
coding configurations as described in [
5
].
The coding efficiency gains in terms of Bjøntegaard Delta (BD) rate [
2
], when
enabling the merge and skip modes, are reported in [
9
]and[
12
] and also summa-
rized in Table
5.2
. Average bit rate savings between 6 % and 8 % are observed. It
can be seen that the gains for the random access main (RA-Main) and low delay
B main (LB-Main) configurations using bi-predictive B pictures are higher than
for the low delay P main (LP-Main) configuration, which is restricted to use uni-
prediction. Since the merge mode only uses a flag and an index to signal all motion
data, it is more efficient the more motion data signaling is omitted that way, e.g.
two sets of motion data in bi-prediction. Another observation is that merge and
skip modes are saving up to 20 % for the class E sequences. These sequences
represent videoconferencing content where the static background can efficiently be
coded using skip and merge modes. More detailed results and a comparison with an
AMVP-based direct mode can be found in [
12
].
