Graphics Reference
In-Depth Information
abrupt changes in the field of motion parameters, which is common in natural video
content. One reason is that the quad-tree structure systematically does not allow for
joint description of child blocks that belong to different parent blocks. Also, the fact
that a block can only be divided into exactly four child blocks will eventually lead
to ineffective borders.
To remedy these inherent drawbacks of the quad-tree structure, HEVC uses block
merging which allows us to code motion parameters very cheaply (in terms of bit
rate) in these ineffective border situations [
9
,
12
,
23
]. The algorithm was inspired
by the work of [
10
], in which the authors show that rate-distortion optimized
tree pruning for quadtree-based motion models can be substantially improved by
introducing a subsequent leaf merging step. In [
22
], the authors study the benefits of
leaf merging on a broader theoretical basis. Here we leave it at the intuitive example
given above and concentrate on the integration of block merging into HEVC.
Following from the observation of ineffective borders, block merging introduces
a terse syntax allowing for a sub-block to explicitly reuse the exact same motion
parameters contained in neighboring blocks. Like AMVP, it compiles a list of
candidate motion parameter tuples by picking from neighboring blocks. Then, an
index is signaled which identifies the candidate to be used. Block merging also
allows for temporal prediction by including into the list a candidate obtained from
previously coded pictures. A more detailed description is given in the following.
5.2.2.2
Merge Candidate List Construction
Although they appear similar, there is one main difference between the AMVP
and the merge candidate list. The AMVP list only contains motion vectors for
one reference list while a merge candidate contains all motion data including
the information whether one or two reference picture lists are used as well as a
reference index and a motion vector for each list. This significantly reduces motion
data signaling overhead. Section
5.2.2.3
describes the signaling in detail as well as
discusses how parsing robustness is achieved. Overall, the merge candidate list is
constructed based on the following candidates:
up to four
spatial merge candidates
that are derived from five spatial neighboring
blocks
one
temporal merge candidate
derived from two temporal, co-located blocks
additional merge candidates
including combined bi-predictive candidates and
zero motion vector candidates
Spatial Candidates
The first candidates in the merge candidate list are the spatial neighbors. Here, the
same neighboring blocks as for the spatial AMVP candidates are considered which
are described in Sect.
5.2.1.1
and illustrated in Fig.
5.4
b. In order to derive a list
of motion vector predictors for AMVP, one MVP is derived from A0 and A1 and
