Page 159 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

136 Chapter 5. Warping-Based Motion Estimation Techniques

At the start of the node-tracking algorithm, the BMA described earlier
was used to provide initial estimates of the inner nodal motion vectors.
Those initial estimates were then re(ned using the iterative procedure of
Sullivan and Baker [108]. In each iteration of this procedure, the nodes
are processed sequentially, where the motion vector of a node is re(ned
using a local search around the motion vector from the previous itera-
tion while holding constant the motion vectors of its surrounding eight
nodes. During this local search, the quality of a candidate motion vector is
measured by calculating the distortion measure between all four patches
connected to the node and their warped predictions from the reference
frame. The local search is applied to a node only if its motion vector, or
the motion vector of at least one of its surrounding nodes, was changed
in the previous iteration. The local search used here examines the eight
nearest candidate displacements centered around the displacement from
the previous iteration. For each frame, 10iterations were used to re(ne
the nodal motion vectors.

During motion estimation and compensation, the bilinear spatial transfor-
mation is employed. This is implemented in the CGI [108] form (de-
scribed in Section 5.2.7), where the motion vector used to compensate
a pel within a patch is bilinearly interpolated, Equation (5.8), from the
four nodal motion vectors at the vertices of the patch.
In BMA-HO and WBA algorithms, bilinear interpolation was used to obtain
intensity values at subpel locations of the reference frame. In each algorithm,
motion was estimated and compensated using original reference frames. Motion
vectors were coded using the median predictor and the VLC table of the H.263
standard. The DFD signal was also transform encoded according to the H.263
standard and a quantization parameter of QP = 10. All quoted results refer to
the luma components of sequences.
Table 5.1 compares the objective prediction quality of the preceding three
algorithms when applied to the three test sequences with a frame skip of 3.
The WBA outperforms the basic BMA by about 0.16 –1:57 dB, depending on
the sequence. However, the WBA fails to outperform the advanced BMA-HO

Table 5.1: Comparison between BMA and WBA in terms of objective prediction quality
Average PSNR (dB) with a frame skip of 3
AKIYO FOREMAN TABLE TENNIS
BMA 39.88 27.81 29.06
WBA 41.45 29.09 29.22
BMA-HO 41.77 29.51 29.87