Game Development Reference
In-Depth Information
Fig. 3.5
Illustration of
neighboring blocks of PU
3.2.3 Motion Vector Prediction and Coding
Motion vector prediction (MVP) plays an important role in inter prediction, which
can reduce the redundancy among motion vectors of neighboring blocks and thus
save large numbers of coding bits for motion vectors. In AVS2, four different pre-
diction methods are adopted, as tabulated in (Table
3.1
). Each of them has its unique
usage. Spatial motion vector prediction is used for spatial derivation of Skip/Direct
mode in F frames and B frames. Temporal motion vector prediction is used for tem-
poral derivation of Skip/Direct mode in B frames. Spatial-temporal combined motion
vector prediction is used for temporal derivation of Skip/Direct mode in F frames.
For other cases, median prediction is used.
In AVS2, the motion vector is in quarter-pixel precision for the luminance compo-
nent, and the subpixel is interpolated with an 8-tap DCT interpolation filter (DCT-IF)
(Lv et al.
2013
). For the chrominance component, the motion vector derived from
luminance with 1/8 pixel precision and a 4-tap DCT-IF is used for subpixel interpola-
tion (Wang et al.
2014
). After motion vector prediction, the motion vector difference
(MVD) is coded in the bitstream. However, redundancy may still exist in MVD, and
to further save coding bits of motion vectors, a progressive motion vector resolution
adaptation method is adopted in AVS2 (Ma et al.
2013
). In this scheme, MVP is
firstly rounded to the nearest integer sample position, and then the MV is rounded to
a half-pixel precision if its distance fromMVP is larger than a threshold. Finally, the
resolution of MVD is decreased to half-pixel precision if it is larger than a threshold.
Table 3.1
MV prediction methods in AVS2
Method Details
Median Using the median MV values of the neighbouring blocks
Spatial Using the MVs of spatial neighbouring blocks
Temporal Using the motion vectors of temporal collocated blocks
Spatial-temporal combined Using the temporal MVP from the collocated block plus an offset
MVD from spatial neighbouring block
