Compression - 3DTV: Processing and Transmission of 3D Video Signals - page 86

Compressed depth map

Step I: Histogram analysis Step II: Bilateral filtering

Segment the image into equal
size segments of 64 × 64 pixels

For each pixel identify the
corresponding segment

Locate the nearest peak

Obtain the histogram of each
segment

Calculate bilateral filter
weights

Smooth the histogram using
an 1-dimensional averaging
filter

Perform adaptive bilateral
filtering

Identify dominant peaks and
their enclosing valleys

Calculate distance to
enclosing valleys for each
peak Recovered depth map

Figure 3.13 Block diagram of the adaptive depth post-processing framework
(taken from [43])

In [44], the authors applied a similar idea in the depth pre-encoding
stage, in order to remove unnecessary spatial variations in the depth map
sequences that would have a not significant influence on the view synthesis
quality. However, by removing the non-required high frequency elements
in the depth map sequences, some useful bit-rate is saved. On the other
hand, significant depth transition areas are preserved by properly adjusting
the filter coefficients. Different from the work presented in [43], the work
reported in [44] uses a colour-video aided tri-lateral Gaussian filter instead
of a bi-lateral filter. This involves an additional filter kernel, which takes
into account the similarity of colour and corresponding depth video edges.
At every depth map pixel to be processed, a window of 2 w

2 w is formed
centred at the corresponding depth pixel. Subsequently, in these kernels of
2 w

×

× 2 w (denoted by ), the filtered depth value is computed as

D p = coef f pq · I q / coef f pq (3.1)

q ∈ q ∈

where, q denotes a pixel within the kernel and p is the centre pixel to be
processed. The coeff is a multiplication of three different factors, namely
the closeness in pixel, similarity in depth value and the similarity in colour