3D Video Representation and Coding - Novel 3D Media Technologies - page 37

Other authors have proposed DC techniques that include specifically designed
intra prediction methods. In [ 37 ], a new prediction method is combined with H.264/
AVC, which defines two regions for the depth-map macroblocks (MB), which are
represented by a losslessly encoded binary or predicted from previously encoded
edge MBs. [ 38 , 39 ] also proposed an additional intra mode for the H.264/AVC
encoder, which adaptively segments each depth MB in order to capture the edge
structure. A constant [ 39 ] or a planar [ 38 ] approximation is then used for each
segment. In [ 37 ] the authors used an approximation of MB segments with a constant
surface for HEVC-based DC.
In spite of the advantages of these edge aware intra prediction modes for DC, the
use of transform coding for compressing the prediction residue often introduces
coding errors in the high frequency components, i.e., the edges. This motivated
the proposal of several techniques which exploit new paradigms for DC. Probably
the most well know is the Platelet encoder [ 40 ]. This algorithm segments the depth
MB using a combination of quadtree decomposition with arbitrary segmentation
using a linear edge, and approximates each segment by a piecewise-linear function.
This efficiently preserves sharp object boundaries and edges, resulting in a high
rendering quality.
Graziosi et al. [ 41 , 42 ] describe the use of an alternative coding paradigm for
DC, based on pattern matching, with state-of-the-art results. The proposedmethod is
referred to as Multidimensional Multiscale Parser (MMP) [ 43 ]. Each MB is first
predicted using a set of prediction modes based on the ones of the H.264/AVC
standard. The resulting residue is then encoded by using approximate patternmatching
with codewords from an adaptive dictionary. The pattern matching step uses blocks of
different sizes (scales), through the use of an appropriate scale transformation, which is
applied to the dictionary elements prior to the block matching.
In [ 36 , 44 ], a DC algorithm that shares some common points with MMP is
presented. The Predictive Depth Coding algorithm (PDC) is based on the
same flexible partitioning used by MMP [ 43 ], combined with an improved predic-
tive framework and a new residue encoding scheme. In [ 36 ] the authors demon-
strate the advantages of using large MB sizes, combined with a very flexible
partitioning scheme. The large MBs are able to efficiently approximate the homo-
geneous regions of the depth-maps. The highly flexible partitioning scheme of PDC
is able to efficiently capture the edge structure of the depth-maps. The result is the
generation of a very low energy residue signal, composed mostly by null samples.
The few non-null residue regions are encoded by simply quantizing the mean value
of the residue and encoding the result with an adaptive arithmetic encoder.
Figure 3.12 represents the rate-distortion results (PSNR vs. the rate required for
DC) for one intermediate view synthesized using the depth maps compressed with
some of the discussed DC algorithms. The intermediate view of virtual camera
4 was synthesized using the compressed depth-maps and the original textures of
views 3 and 5 of the well-known Ballet (left) and Breakdancers (right) sequences.
Results for the Platelet [ 40 ], MMP [ 42 ], and PDC encoder [ 36 ] are compared with
the state-of-the-art, transform-based H.264/AVC and H.265/HEVC for the intra
compression of depth-maps. Both plots show the consistent advantage of PDC
algorithm when compared with other state-of-the-art depth map coding algorithms.