Graphics Reference
In-Depth Information
CUs, the RQT subdivision is independent of the PU partitioning, which may lead
to a transform block covering several different PBs having the same CB as the
root. Note that even though applying the transform across PB boundaries can cause
unwanted coding efficiency degradations for several CBs, it may also increase the
coding efficiency for other CBs. Experimental results [
23
] showed that giving the
encoder the freedom to decide whether a transform is applied across PB boundaries
increases the average coding efficiency by about 0.4-0.7 % Bjøntegaard Delta bit
rate (BD rate) [
3
].
As discussed in Sect.
3.2.3
, for intra-predicted CUs, the splitting of the luma and
chroma CBs into transform blocks does not only determine the size of the transform
that is applied for residual coding, but also the size of the blocks for which a single
intra prediction signal is generated. If the size of a luma CB size is equal to the
minimum CB size signaled in the sequence parameter set, it is possible to signal
four luma intra prediction modes for the luma CB, each for one of the four equally-
sized square subblocks. In that case, a transform block cannot span the complete
luma CB, since the neighboring subblocks within the luma CB that precede a current
subblock in coding order have to be fully reconstructed for generating the prediction
signal for the current subblock. As a consequence, if a luma CB is subdivided into
four subblocks for transmitting intra prediction modes, it is also subdivided into four
TBs. Note, however, that each of the resulting TBs may then be further subdivided.
A leaf node is implied in the RQT when the minimum transform size is reached,
i.e., if the binary logarithm n of the actual transform block size satisfies the condition
n
D
n
min
. However, the recursive subdivision is also restricted by the maximum
RQT depth, even when smaller TBs are allowed by the minimum transform size,
which corresponds to the case that the actual RQT depth d conforms with d
D
d
max
and that the above-mentioned conditions on implicit subdivision are not fulfilled.
Note that at most one single transform tree syntax is transmitted for each CU, and
thus, the same RQT structure determines the TBs of all color components, resulting
in the same subdivision of luma and chroma components. An exception to this rule,
however, is given for the case of a 4
4 luma TB in a 4:2:0 chroma-formatted video
signal, where the subdivision would lead to a corresponding 2
2 chroma TB, which
is not supported in HEVC. Therefore, an RQT is allowed to split an 8
8lumaTB,
but not the corresponding 4
4 chroma TB, leading to a different interpretation of
the corresponding RQT structure for luma and chroma in this particular case. A
transform unit (TU) represents a luma TB greater than 4
4 samples or four luma
TBs with a size of 4
4 samples, the corresponding two chroma TBs, and the
associated syntax structures.
3.2.4.2
Parameter Signaling
The RQT parameters introduced in the previous subsection, i.e., the maximum RQT
depth d
max
and the binary logarithms of the minimum and maximum transform sizes
n
min
and n
max
are transmitted in the bitstream at the SPS level. Regarding the RQT
depth, different values can be specified and signaled for intra- and inter-predicted
