Image Processing Reference
In-Depth Information
as autostereoscopic displays, requires simultaneous decoding of a number of views.
It follows that the decoder should be capable of rendering as many views as possible
to provide an immersive 3D experience. To satisfy these needs, view synthesis can
be employed at the receiver to generate the necessary views by interpolation from
the actually received views. A commonly used synthesis algorithm is depth-image-
based rendering (DIBR) [
27
], where the depth maps are used to provide the
necessary geometry during the interpolation process. Depth estimation is a high
computational task, which cannot be performed at the decoder in real-time appli-
cations [
28
] and, as a result, has to be transmitted together with the texture video.
This gives rise to the video-plus-depth coding schemes discussed in this section.
3.3.3.1 Depth Map Coding Techniques
The search for standardization of efficient compression techniques for MVD is
being conducted by the Joint Collaborative Team on 3D video coding extension
development (JCT-3 V). JCT-3 V is composed by elements from MPEG
(ISO/IECJTC1/SC29/WG11 3DV ad hoc group) and ITU-T (SG 16 WP 3). The
compression of MVD is in line with most previous video coding standardization
efforts and most of the current proposals are being developed in a H.264/AVC or
H.265/HEVC framework. Nevertheless, several proposals for depth-map coding
(DC) techniques have been made outside of the scope of standardization groups. In
this section we will review the most relevant techniques for DC.
Depth-maps are bi-dimensional matrixes of elements that represent the depth
associated with each pixel of the associated texture view. For this reason, DC is
often regarded as an image compression problem. Nevertheless, efficient DC
techniques must take into account the particular features of depth-maps. Depth-
maps are commonly composed by large homogeneous areas, associated with the
objects at a given depth-plane, surrounded by sharp edges, corresponding to the
objects
boundaries. Since depth-maps will be used for DIBR, and not for display,
the compression efficiency for DC should be measured based on the distortion of
the synthesized intermediate views. DIBR is particularly sensitive to errors in the
edges of the compressed depth-maps. These edges correspond to high frequency
regions, which are strongly affected by traditional transform-based encoders, and
should be taken into account during the development of DC.
The MPEG standardization groups are working on AVC and HEVC-based
encoders for depth-maps. For AVC, the encoder is referred to as 3D-AVC
[
29
]. The extension to the HEVC standard supporting DC is known as 3D-HEVC
[
30
,
31
]. Both these encoders incorporate specific tools for the independent com-
pression of depth-maps or for the joint compression of texture-plus-depth.
The 3D-HEVC techniques for DC are based on the traditional hybrid video
encoding model, which uses predictive coding combined with transforms and
entropy encoding [
12
]. 3D-HEVC uses specific encoding setups (e.g., disabling
the de-blocking filter and the sample adaptive offset filter), but new tools for DC
have also been developed, including: depth modeling modes, region boundary
'
