Graphics Reference
In-Depth Information
Table 11.3
Bit estimation algorithm comparison
Fast
RDO
Algorithm
JCTVC-G763(HM)
CFBAC
Bypass-based
only
BD-Rate[%] (vs. CABAC)
0:13
1.14
2.65
48.31
Fig. 11.17
CFBAC
architecture block diagram
State Updater
Final Mode
Decision
State Memory
Context
Modeler
State Bits
LUT
*
Binarization
0/1 Bit
Counter
*
CFBAC
state memory between modes. The issue for sharing state memory is that states are
updated in arithmetic coding process. To eliminate this issue, we use fixed state
memory that is not updated during bit estimation. However, if the states used in
the bit estimation are too different from the actual states, the bit estimation will not
be accurate enough and cause low quality decision in HCMD. Hence, we keep the
context fixed at a CTU-level. bit rate increase for this scheme is limited [
17
]. The
states are the same in CABAC at the beginning of the CTU. During the bit estimation
process inside CTU, the states are not updated. After the final mode decision is
made, the bits for the selected mode are traversed and the final states are updated.
For more simplification, we also uses bit estimation table in JCTVC-G763 [
1
]for
bits look-up instead of arithmetic encoder in context-fixed scheme.
For quality comparison, the BD-rate differences vs. CABAC-based bit estimation
are shown in Table
11.3
. HM takes JCTVC-G763 as the default fast bit estimator.
As we can see, the quality drop would be high if all the mode decision is done only
by fast RDO. For bypass-based method, the quality loss is moderate and hardware
cost is low. If more accurate result is required, CTU-based CFBAC can be used.
The hardware architecture for CFBAC is shown in Fig.
11.17
. Since the states
are fixed, all the MPS and LPS coded using the same type of context share the same
probability. For every '1' bin and '0' bin in the same context, the bits produced is a
fixed number B
0
and B
1
according to the JCTVC-G763 look-up table, respectively.
So we need only to count the number of '1' bins and '0' bins for bit rate estimation.
We modify the binarization process and produce the 1's count C
1
and 0's count C
0
.
The bit rate can be estimated according to Eq. (
11.1
).
X
F
bits
.i np
u
t/
D
B
0
.n/
C
0
.n/
C
B
1
.n/
C
1
.n/
(11.1)
8n2contexts
