Graphics Reference
In-Depth Information
3.2.4.3
Fast Encoder Control
The number of different RQT partitionings as well as the number of different coding
tree partitionings grows faster than the double exponential expression 2
4
d 1
with
increasing depth d of the tree. However, as further explained in [
25
], by application
of the generalized BFOS algorithm [
5
], the derivation of an optimal partitioning in
a rate-distortion sense can be achieved without the need of a brute-force exhaustive
search that requires to compare each partitioning option with all of its competing
options. In fact, it can be shown that without applying any early termination strategy,
the computational complexity of the fast tree-pruning process is proportional to
.4
d
1/=3, which is the number of internal nodes of a quadtree with maximum
depth d . However, in order to further reduce the computational complexity for the
tree-growing process at the encoder side, heuristic early-pruning techniques [
25
,
38
]
can be additionally applied.
In case of the RQT partitioning, such an algorithm was presented in [
38
]. Here,
the idea is that the evaluation of further subdivisions at a given RQT node should
be terminated when all magnitudes of unquantized transform coefficients are below
an appropriately chosen, quantizer step size-dependent threshold. It was shown that
by applying such a strategy, the encoder runtime can be reduced by about 5-15 %
with only minor impact on coding efficiency. Also, the reduction of encoder runtime
is consistently higher for a larger maximum RQT depth, since with a larger search
space for the optimal RQT partitioning, typically larger improvements in complexity
reduction can be achieved if only a subset of the whole search space is considered.
For more details, the reader is referred to [
38
].
3.2.5
Performance
For evaluating selected design aspects of the HEVC block structures, we performed
coding experiments for two different application scenarios. The first scenario
considers the coding of high-resolution video with entertainment quality, while the
second scenario addresses interactive video applications such as video conferencing.
For the scenario of entertainment applications, we selected five HD sequences of
varying content, all having a resolution of 1920
1080 luma samples. The sequences
were coded using a dyadic high-delay hierarchical prediction structure with groups
of eight pictures [
36
] and four active reference pictures. Random access points,
which are coded as I pictures, are inserted in regular intervals of about 1 s. In order
to enable clean random access, pictures that follow an I picture in both coding and
display order are restricted in a way that they do not reference any pictures that
precede the I picture in coding or display order. With exception of the random access
pictures, all pictures are coded as B pictures. The quantization parameter (QP) is
increased by 1 from one hierarchy level to the next and the QP for the B pictures of
the lowest hierarchy level is increased by 1 relative to that of the random access I
pictures. All pictures are coded as a single slice.
