Page 157 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

134 Chapter 5. Warping-Based Motion Estimation Techniques

or to interpolate the motion (eld (as described in Section 5.2.7) before being
able to perform motion compensation.

5.3 E1ciency of Warping-Based Methods at Very
Low Bit Rates

This section investigates the performance of warping-based methods and com-
pares it to that of block-matching methods. The main aim is to answer the
following question: Are there any gains for using higher-order motion models
at very low bit rates? In other words, this section assesses the suitability of
warping-based methods for applications like mobile video communication.
Most results reported in the literature compare a warping-based algorithm
to the basic block-matching algorithm. The authors feel that this is an unfair
comparison for the following reasons:
1. As shown in Section 5.2.7, in warping-based compensation the motion
vector used to compensate a pel in a given patch is interpolated from
the nodal motion vectors at the vertices of the patch. Although the nodal
motion vectors may be at full-pel accuracy, the resulting interpolated
motion vector is at subpel accuracy. It is unfair to compare this subpel
compensation to the full-pel compensation of the basic block-matching
algorithm. A more fair comparison would be with a subpel (at least
half-pel) block-matching algorithm.
2. Again, from Section 5.2.7, a warping-based method calculates one
motion vector per pel. Thus, each pel within a patch is compensated
individually. It is unfair to compare this to the basic block-matching al-
gorithm, where the whole block is compensated using the same motion
vector. A fairer comparison would be with overlapped motion compen-
sation, where each pel within the block is compensated individually, as
evident from Equation (4:32).

3. A warping-based method is much more computationally complex than
the basic block-matching method (as is shown later). This increased
complexity gives the warping-based method an unfair advantage over
the basic block-matching method. To provide a fairer comparison, the
basic block-matching method must be augmented by some advanced
techniques (like subpel accuracy and overlapped compensation).
Thus, in this study, the following algorithms were implemented:
BMA This is a full-search full-pel block-matching algorithm with 16 × 16
blocks, restricted motion vectors, a maximum displacement of ± 15 pels,
and SAD as the matching criterion.