Page 117 - ITU Journal, ICT Discoveries, Volume 3, No. 1, June 2020 Special issue: The future of video and immersive media

ITU Journal: ICT Discoveries, Vol. 3(1), June 2020

tiparty framework and, therefore, the JPLM design should 4.5.2 Objective quality assessment
favour successful bug-fix actions, allow the addition of
The JPEG Committee measures the performance of propo-
new capabilities or functionality, and enable the aggrega-
nent codecs against the anchors using a variety of objec-
tion of subsystems (in case of creation of new features via
tive measures including bit rate, and for random access,
future amendments, for instance). In other words, JPLM
the ratio of bits decoded to access a region of interest ver-
architecturehasbeendesignedhavinginmindfuturesup-
sus bits required to decode the entire point cloud. How-
port for point cloud and holography extensions.
ever, here we will expand on the measures of geometric
fidelity employed by the JPEG Committee. The measures
JPLM complies with JPEG’s recommended practices for
described here have been chosen to be consistent with
software development and maintenance.This compliance
the MPEG Committee, in line with the desire of the JPEG
requires a permissive software license and, therefore,
Committee to be consistent with other efforts in this field.
JPLM uses a BSD 3-clause license. Moreover, JPLM imple-
These are as follows:
ments unit tests extensively. These tests ensure that im-
plemented functionalities remain during software devel- Average Point to Point Distance (D1): This measure starts
opment. These tests also form part of the development with the identification of corresponding points on a ref-
pipeline that adopts continuous integration (CI) tools. erence and processed point cloud by a nearest neighbour
algorithm. The Euclidean distance between the corre-
4.5 JPEG Pleno point cloud sponding points on the reference and processed point
clouds is then calculated. The set of distances computed
4.5.1 Current status in this way for the entire processed point cloud are then
arithmetically averaged as the quality measure [21].
Support for point cloud coding will form a future part of
Average Point to Plane Distance (D2): For this measure,
the JPEG Pleno standard and the JPEG Committee is cur-
a nearest neighbour algorithm identifies corresponding
rently working towards this goal. As a necessary first step
points between the reference and processed point cloud.
in this direction, the JPEG Committee is currently work-
A tangent plane is then fitted to the neighbourhood of the
ing on determining the best objective and subjective as-
point on the reference cloud taking into account any nor-
sessment methodologies for this activity. This will sup-
mal information available and the normal of the plane is
port the later evaluation of point cloud coding solutions
used to compute the smallest distance between the point
to be added to the JPEG Pleno standard.
on the processed cloud and the plane fitted to the refer-
ence cloud. The set of distances computed for the entire
The JPEG Committee plans to compare different compet-
processed point cloud are then arithmetically averaged as
ing coding solutions against anchor point cloud codecs
the quality measure [21].
used with well-defined parameter settings to reference
encoded contents at a variety of quality levels. To be con- Average Plane to Plane Similarity: This metric is based
sistent the state of the art for point cloud coding, the V- on the angular similarity of tangent planes of corre-
PCC and G-PCC codecs in use by the MPEG Committee sponding points between the reference and processed
have been selected as anchor codecs [21]. point clouds. A nearest neighbourhood algorithm identi-
fies corresponding points in the reference and processed
Point cloud quality assessment is essential to selecting
point clouds and tangent planes are fitted to the local ar-
and evaluating potential submissions to form part of a
eas of the corresponding points on both the reference and
new standard in this area. Point positions may be ar-
processed point clouds. The minimum angle between the
ranged on a regular grid (for example on a voxel grid) or
two planes is then computed and used to compute an er-
irregularly arranged in 3D space. In addition to geometri-
ror value between the tangent planes [23] for every point
cal information, attributes such as colour, surface normal,
on the processed point cloud. The mean square of the er-
texture and bump maps or scientific data such as tem-
ror values is then computed and the logarithm of a ratio of
perature may be associated with individual points or the
a constant and this mean square value is used as the error
point cloud as a whole. The manner in which the point
measure (akin to a peak signal to noise ratio).
cloud is rendered, including whether the points are ren-
dered as individual points or are connected by a mesh and Point to Point Attribute Measure: This measure starts with
how the mesh is formed can also greatly change subjec- the identification of corresponding points on a reference
tive judgments of quality [22]. Point cloud data is 3D in and processed point cloud by a nearest neighbour algo-
nature, however current 3D display systems such as 3D rithm in line with the D1 metric. For colour attributes,
monitors or Head-Mounted Displays (HMDs) often have the MSE for each of the three colour components between
lower resolution and narrower colour gamut than high- the corresponding point is calculated in YCbCr space. The
level 2D monitors typically used for subjective quality set of colour distances computed in this way for the entire
assessment. This further complicates subjective assess- processed point cloud are then arithmetically averaged as
ment of quality. a final PSNR quality measure [21].

© International Telecommunication Union, 2020 95