Page 188 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

Section 7.5. Techniques Based on a Subsampled Block-Motion Field 165

In other words, it is very common to )nd neighboring blocks with identical or
nearly identical motion vectors. Thus, in this category, a subsampled block-
motion )eld is )rst obtained by estimating the motion vectors for a fraction
of the blocks in the frame. This )eld is then appropriately interpolated to
determine the motion vectors of the remaining blocks.
Liu and Zaccarin [157] use a checkerboard subsampling pattern and estimate
the motion vectors for half of the blocks (i.e., a 2:1 subsampled motion )eld)
using full search. Then they estimate the motion vectors of the other half
using a limited search that examines only four candidate motion vectors. Those
candidates are the four surrounding motion vectors that were estimated using
full search. For example, in Figure 7.3(a) the motion vectors of blocks B, C,
D, and E are estimated using full search. Only those four vectors are then used
as candidates when estimating the motion of block A. This algorithm reduces
complexity by roughly a factor of 2, with only a slight loss in prediction
quality.
Another algorithm was also proposed by Liu and Zaccarin in Ref. 157.
In this algorithm, each block is divided into four subblocks. Motion is )rst
estimated, using full search, for one subblock in each block, say, the top-left
subblock. The motion vectors of the remaining subblocks are then estimated
using a limited search with candidates from the neighboring full-search motion
vectors. For example, in Figure 7.3(b) the motion vectors of subblocks A, B,
C, and D are estimated using full search. Only those four vectors are then
used as candidates when estimating the motion vectors of subblocks a, b,

N

N
N B
N/2
N/2 A a B
E A C N
b c

C D
D

(a) Subsampling with blocks (b) Subsampling with subblocks
Figure 7.3: Reduced-complexity using the subsampled motion )elds of Liu and Zaccarin [157]