Image Processing Reference
In-Depth Information
Fig. 6.11 2D (
left
) and 3D (
right
) heatmap
6.3.2.1 Conclusion
It was observed that heat maps obtained from the eye tracker almost matches with
the salient areas detected by the TIS approach [
29
]. It is also observed that face
plays an important role and it indeed grabs visual attention wherever it is present as
also shown in Fig.
6.10
that reveals multiple users points of focus with the strong
emphasis on face. Additionally the captured data (2D and 3D sequences) using eye
tracker were compared. No big differences in distribution of gaze fixation positions
were found here so in general it can be assumed that the viewing behavior does not
differ much. There were some minor hints though where it might be assumed that in
specific situations or properties of the clips a viewer might tend to watch the
foreground more intensive in 3D than in 2D as shown in Fig.
6.11
.
6.3.3 VAM Based Differentiated Bit Allocation
for Video Coding
In a general context and without specific requirements, a saliency map is classically
provided as a dense gray level image (8 bits per pixel). As described in Sect.
6.2.5
,
for some post-processing purposes (e.g., video coding), the saliency regions can be
merged into a few number of regions with different gray levels, which correspond
to the level of interest of the region. Especially for video compression, this image
partitioning is very useful in order to optimize the bit rate allocation by decreasing
the bit rate budget for less salient regions.
In a lossy video coding process, the quantization factor (Qp) is the parameter that
allows to allocate more or less bit rate, regarding a target quality level. Especially,
in H.264/AVC [
38
], the quantization parameter is set frame by frame and specific
offsets can be applied at the macroblock (16
16 pixels) granularity. In this
context, a dense disparity map is no longer necessary and the map is downscaled
before being passed to the encoder (see Figs.
6.7
d and
6.8
d). Once this quantized
and downscaled disparity map is obtained, the saliency levels are mapped to the
corresponding offsets. In order to improve efficiently the encoding process, the Qp
offset mapping must be done by smoothing the regions where saliency variations
