Information Technology Reference
In-Depth Information
bi-directional motion estimation, [8], or an adaptive combination of all three search
directions, [10].
As an integral part of the motion estimation, search patterns with fast convergence
rates and capability of taking into account
true
object motion should be utilized
in order to interpolate more pleasant inter-frames. Among many algorithms, three
step[14], spiral [6], hexagonal [15], diamond and modified diamond, [16] - [17], are
the most popular search patterns utilized for FRUC. Although, in general, it fails to
estimate correct motion vectors, [13], several techniques utilize
full-search
method,
[2] - [8].
Mean absolute differences (MAD)
([6, 15, 16]) and
sum of absolute differences
(SAD)
([4, 9, 11, 13]) are widely exploited match error criteria in block-based mo-
tion estimation. While SAD avoids division operation, and hence provides com-
putational efficiency over MAD, it does not enforce spatial smoothness. In order
to alleviate this problem, Choi et. al, [3], combine SAD measure with side match
distortion as a way to estimate relatively smooth motion fields.
The main drawback of block based motion estimation and compensation ap-
proaches is that they suffer from blocking artifacts especially occurring at object
boundaries. As an attempt to reduce these artifacts, Ha et al., [2], utilize larger blocks
for motion estimation, whereas the block size is reduced during motion compensa-
tion. In a similar spirit, there are also pixel-based approaches, [20, 21] , which con-
vert block-based motion fields to pixel-based motion fields prior to compensation.
Similarly, [4] adaptively changes motion compensation block size depending on the
size of the objects. Alternatively, a class of methods cope with blocking artifacts
by using overlapping blocks in motion compensation stage [8, 3]. With the same
concern, Lee et al. consider neighboring blocks' motion trajectories by computing
a weighted average of multiple compensations [19].
Most of the algorithms in the literature employ a two-frame approach for motion
estimation and compensation. Although the utilization of two frames may not ef-
fectively handle the occlusion problem , such an approach is enforced by existing
hardware limitations, caused by the adversity of providing extra storage for multi-
frames. However, there exist algorithms that exploit multiple frames at the cost of
increased complexity [9, 10].
Techniques that are based on translational motion model (such as the ones men-
tioned above) generally rely on the idea of linearly interpolating motion vectors
between the original successive frames [22] - [26]. Alternatively, a higher order
function, such as a polynomial, can be used for the interpolation as proposed in [27].
On the other hand, the interpolation problem becomes more complicated for higher
order parametric motion models (e.g., six-parameter affine and eight-parameter ho-
mography mappings) due to the need for defining a reasonable model parameter
space.
Although blocking artifacts inherent in block-based techniques are intended to
be solved by using previously mentioned techniques, they cannot be avoided com-
pletely at object boundaries, where multiple-motion exists. In an attempt to segment
true object boundaries, region-based FRUC techniques estimate arbitrarily-shaped
