Full Parallax 3D Video Content Compression - Novel 3D Media Technologies - page 59

also proposed, in which a sparse integral image is reconstructed from the set of
views of the layer 1 and is inserted in the reference picture buffer to encode layer 2.
The wavelet-based and 3D-DCT-based methods are conceptually far from the
standard encoders schemes (H.264/AVC and HEVC) and more adequate for still
image coding, whereas the self-similarity and multi-view methods are more easily
included in the structure of these reference video encoders.

4.3.1.2 Multi-View Video Coding Standards

Both H.264/AVC and HEVC standards have multi-view extensions [ 34 ]: respec-
tively MVC and MV-HEVC, which provide additional high level syntax that
enables the inter-view prediction. 3D-HEVC is an extension for the multi-view
plus depth (MVD) format, which provides tools related to depth maps, new tools at
the Coding Unit level (CUs in HEVC replace H.264/AVC macroblocks) for side
views and new inter-component dependencies. These standard multi-view encoders
are designed to handle horizontal parallax content with a small number of views.
The Neighboring Block Disparity Vector (NBDV) [ 44 ] and the Inter-View
Motion Prediction (IVMP) [ 45 ] are specific 3D-HEVC coding tools. In the
HTM7.0 reference software of the standard (used for experiments in the following)
they are implemented as follows. NBDV searches through already coded temporal
and spatial neighboring CUs for a disparity vector (DV) and derives it for the
current CU (by simple copy). Motion vectors and temporal references (motion
parameters) of the CU pointed by this DV in the reference view are used by IVMP
to create the Inter-view Predicted Motion Candidate (IPMC). IPMC is introduced at
the first place in the merge [ 46 ] candidate list, and the DV itself is inserted in the
merge list as Disparity Motion Vector candidate (DMV). This description corre-
sponds to the implementation of the HTM7.0 reference software, and modifications
of these steps can occur during the further evolutions of the standard.

4.3.1.3 Improvements for Full Parallax

Full parallax SMV content can be encoded with a multi-view standard encoder with
an adaptation of the inter-view references structure level. In [ 33 ], the views are first
scanned in spiral as illustrated in Fig. 4.12a and realigned horizontally. The
horizontal arrangement is then encoded by MVC with an IBP prediction structure
(Fig. 4.12b ). The resulting scheme of equivalent IBP structure with the views in two
dimensions is illustrated in Fig. 4.12c , which shows that this approach is limited by
the introduction of unsuitable and ineffective predictions.
In [ 47 ], horizontal IPP or IBP structures (Fig. 4.13e, f ) are applied to each line of
the views array, and a vertical inter-view prediction is added for the first or central
column of views only, as illustrated in Fig. 4.14a-c . The main drawback of such
structures is the limited number of available vertical inter-view predictions.