Graphics Reference
In-Depth Information
a single PU, the .M=2/
.M=2/ mode is only supported for the minimum CU
size that is signaled in the sequence parameter set. For splitting a CU into two
PUs, HEVC supports six partitioning modes. In the M
.M=2/ mode, the CU
is vertically subdivided into two rectangular PUs of the same size. Similarly, the
.M=2/
M mode horizontally subdivides the CU in two PUs of the same size.
In addition to these symmetric partitioning modes, four asymmetric partitioning
modes are supported, which subdivide the CU into two rectangular PUs of different
sizes, as is illustrated in the bottom row of Fig.
3.6
. One of the resulting PUs has a
rectangular shape with one side having a length equal to the width and height M of
the CU and the other side having a length equal to M=4; the other PU covers the
remaining rectangular area of the CU. The asymmetric partitioning modes are only
supported for CU sizes larger than 8
8 luma samples. Furthermore, for minimizing
the worst-case memory bandwidth, the .M=2/
.M=2/ mode is only supported if
the selected minimum CU size is larger than 8
8 luma samples, so that blocks of
8
4 and 4
8 are the smallest block sizes that can be used for motion-compensated
prediction. In addition, PUs of 8
4 and 4
8 luma samples are restricted to use a
single motion hypothesis.
Supporting more modes for partitioning a picture block into subblocks used
for motion-compensated prediction typically provides the potential for increasing
the coding efficiency. However, this potential can only be exploited if an encoder
evaluates a significant number of the supported partitioning modes. Otherwise,
the syntax overhead associated with an increased number of modes may actually
decrease the coding efficiency in comparison to supporting a smaller set of modes.
The set of partitioning modes supported in HEVC has been selected as a reasonable
trade-off. In addition, HEVC provides the possibility to disable the asymmetric
partitioning modes via a syntax element coded in the sequence parameter set (SPS).
If these modes are disabled, the entropy coding of the partitioning mode is modified
so that the remaining partitioning modes can be signaled with less bits (see Chap.
8
).
This feature is particularly useful for low-complexity encoders that do not have the
computational resources to evaluate all possible partitioning modes.
In the development of video coding standards, one key aspect for improving the
coding efficiency from one generation of standards to the next was to increase
the number of supported block sizes for motion-compensated prediction. If we
concentrate on profiles intended for the coding of progressive video, H.262
j
MPEG-2 Video supports only a single block size of 16
16 luma samples for
specifying motion parameters. In H.263 and MPEG-4 Visual, a subdivision of the
16
16 macroblocks into four 8
8 blocks is additionally allowed. The concept of
subdividing a 16
16 macroblock into smaller blocks is further extended in H.264
j
MPEG-4 AVC, where block sizes from 4
4 to 16
16 luma samples, including non-
square blocks, are supported. Now in HEVC, the concept is additionally extended
towards larger block sizes, resulting in motion compensation block sizes that range
from 4
8 and 8
4 luma samples to 64
64 luma samples. The motion compensation
block sizes that are supported in different standards (excluding special tools for
interlaced video) are summarized in Table
3.1
.
