Game Development Reference
In-Depth Information
Firstly, initialize the variable of reference related. The current random access point
(
currRAP
) and the last reconstructed picture are set 0. And decode sequence header
and then set next random access point (
nextRAP
).
Secondly, derive the actual COI. Decode the syntax element coi in picture header.
And if the coi is smaller than the coi of last decoded picture, add 256 to current COI.
And the actual COI of current picture equals,
COI
=
coi
+
COI
_
Offset
(9.14)
Thirdly, derive theRCS of current picture. Decode the syntax element
use
_
rcs
_
flag
from the header of current picture. And if
use
_
rcs
_
flag
is 1 then decode the
rcs
_
index
and find the RCS that the sequence header points at. Otherwise, decode the RCS from
the picture header.
Fourthly, derive the POI of current decoding picture. The
delta
_
poi
is derived
from the RCS. And the actual POI is derived as follows:
POI
=
COI
−
delta
_
poi
(9.15)
The
currRAP
will be set as the value of
nextRAP
if POI is larger than the
currRAP
.
Fifthly, set up reference picture list. The pictures in the buffer that satisfy Eq. (
9.12
)
are selected as reference pictures of current coding picture.
Finally, derive useless picture information from buffer and delete from buffer.
A picture in buffer is considered useless if it is not referenced by other pictures
and needed to be removed from DPB. Firstly, set the number of useless pictures
num
_
of
_
useless
, which is derived from RCS of current picture. Secondly, derive the
COI of useless reference picture.
COI
_
useless
=
COI
−
delta
_
coi
_
useless
(9.16)
where
COI
_
useless
represents the coding order of useless picture in buffer. The
variable
delta
_
coi
_
useless
could be derived from RCS.
9.1.2.4 Hierarchical P Coding Structure
Based on the flexible reference picture management scheme, a hierarchical P coding
structure is designed and configured to improve coding efficiency. Traditionally, the
QP of P pictures is set the same and four nearest previous constructed pictures are
selected as reference pictures. In the proposed method, four pictures are configured
as a GOP and pictures of different positions of a GOP are coded with different QP.
As can be seen from Fig.
9.2
, the picture that signaled with the black color uses the
smallest QP and is referred the most times. The results are as listed in Table
9.2
.
Compared with the RD4.0 anchor, it can be seen that the performance gain of the
proposed hierarchical P can be up to 9.6, 9.5, and 9.7% on Y, U and V, respectively.
And the average performance gain is 4.3, 7.0 and 7.3%, respectively.
