This commit aims to reduce te amount of copy+paste code used to
support the three different geometry coding modes by unifying them
into a single flow.

This commit adds a DecoderParameters struct to reduce the overhead of
adding extra parameters.

This commit introduces signalling for geometryCodecType in the bitstream
headers in order to allow the decoder to be more easily slaved to the
bitstream -- ie, to avoid needing to configure a special decoding mode.

This commit aims to reduce the overload on the --mode option and to
restore it to the previous encode/decode functionality.  To this end,
a new option (--geometryCodec=...) is added that selects the geometry
codec.

Version v1.0 use of --mode=... is supported.

This commit adds a helper to add / remove colour and reflectance
attributes in a single shot.

This commit introduces PCCTMC3Encoder3Parameters::getAttrParams(...),
a helper to reduce the clutter in searching for attributes to code.

Little is gained from the return type checks, it mostly serves to
clutter the code.

Philip A. Chou authored Mar 29, 2018 and David Flynn committed Jun 05, 2018

This commit does not add the triangle soup geometry compressor to TMC3,
rather.  Rather, it adds a means to call out to TMC1 binaries to perform
the compression, integrate the results into the bitstream.

Attribute compression is performed natively by the TMC13 codec.

Rather than coding all syntax elements for both transform types
irrespective of the in-use type, this commit codes only those that
are necessary for a given transform type.

Philip A. Chou authored Apr 03, 2018 and David Flynn committed Jun 05, 2018

This commit introduces the rahtLeafDecimationDepth parameter to allow
chroma sub-sampling by decimating the coefficients in the bottom n
levels of the tree.

The default value of 3 provides chroma sub-sampling by a factor of 2
in each dimension.

Philip A. Chou authored Mar 25, 2018 and David Flynn committed Jun 05, 2018

This commit ports the Region Adaptive Hierarchical Transform for
attribute coding with colour and reflectance variants.

Khaled Mammou authored Mar 25, 2018 and David Flynn committed Jun 05, 2018

This commit reorganises the attribute coders to allow integration
of the RAHT variant.

In order to support multiple different attribute compression schemes,
this commit moves the existing attribute compression code into a
self-contained class.

In order to support multiple different attribute compression schemes,
this commit moves the existing attribute compression code into a
self-contained class.

Since the generated config files include a reference to the source
file, this commit also regenerates the config.

In order to avoid obscure broken builds when switching branches
or otherwise adding/removing files, this commit removes the use
of wild cards for source files in the tmc3 directory.

To add files to or remove from the build, the tmc3/CMakeLists.txt
must be modified, thereby permitting the build system to detect
changes in the file list.

scripts/Makefile.tmc13-step provides example usage on how to run the codec
to encode an input ply file.

To update the generated config files, run the following:

  $ cd cfg
  $ ../scripts/gen-cfg.sh

Khaled Mammou authored Mar 25, 2018 and David Flynn committed May 17, 2018

When coding attribute prediction deltas, this commit implements a scheme
to code the first and subsequent attribute components using each:

 - one 65-ary contexts to code a delta prefix
 - one binary context to code remainder values

This replaces the previous scheme using:

 - one n-ary (max 2048) context to code a delta prefix (first comp.)
 - one binary context to code remainder values         (first comp.)
 - 512 n-ary (max 2048) contexts to code a delta prefix (subsequent comp.)
 - 512 binary contexts to code remainder values         (subsequent comp.)

Extracted-from: http://wg11.sc29.org/svn/repos/MPEG-I/Part5-PointCloudCompression/CE/CE0/CE0.1/trunk@1357


Reintegrated-by:  <dflynn@blackberry.com>

Khaled Mammou authored Mar 25, 2018 and David Flynn committed May 17, 2018

This commit uses a constant (per attribute) number of neighbours in
point attribute prediction, replacing a method where the encoder searches
for an optimal number and signals this on a per point basis.

Extracted-from: http://wg11.sc29.org/svn/repos/MPEG-I/Part5-PointCloudCompression/CE/CE0/CE0.1/trunk@1357


Reintegrated-by:  <dflynn@blackberry.com>

Khaled Mammou authored Mar 25, 2018 and David Flynn committed May 17, 2018

This commit reduces the number of nearest neighbours used in attribute
prediction to 4 from 8 in order to reduce memory usage.

Extracted-from: http://wg11.sc29.org/svn/repos/MPEG-I/Part5-PointCloudCompression/CE/CE0/CE0.1/trunk@1357


Reintegrated-by:  <dflynn@blackberry.com>

Khaled Mammou authored Mar 25, 2018 and David Flynn committed May 17, 2018

This commit removes the ability to control the quantization dead-zone
parameters since they have always been identical to the step size.

Rather than hard-coding the dead-zone size parameter to be the same as
the step size, this commit:

 - Sets the dead-zone to be ±(2 * stepsize / 3) rather than ±stepsize

 - Does not narrow the interval following the dead-zone.

Extracted-from: http://wg11.sc29.org/svn/repos/MPEG-I/Part5-PointCloudCompression/CE/CE0/CE0.1/trunk@1357


Reintegrated-by:  <dflynn@blackberry.com>

Khaled Mammou authored Mar 25, 2018 and David Flynn committed May 17, 2018

This commit fixes a syntax issue where attribute quantisation step sizes
greater than 2^{16}-1 could not be correctly represented.

Extracted-from: http://wg11.sc29.org/svn/repos/MPEG-I/Part5-PointCloudCompression/CE/CE0/CE0.1/trunk@1357


Reintegrated-by:  <dflynn@blackberry.com>

Ohji Nakagami authored Apr 13, 2018 and David Flynn committed May 11, 2018

This commit eliminates the blending of point attribute values during
the attribute transfer process used in geometry quantisation by forcing
the blending weight to be 0|1 (m42141).

m42538 also recommends to disable the bounded search of the closest
attribute from the contributing set by setting searchRange=0.  This is
reflected by a change to the software defaults and configuration.

Ohji Nakagami authored Feb 19, 2018 and David Flynn committed May 11, 2018

This commit fixes an issue with source sequences that contain 16-bit
reflectance data, which may be clipped to 8-bit ranges during attribute
transfer (recolouring) following geometry quantisation.

This commit provides an implementation of m42239, permitting early
coding termination of sub-trees that contain few points:

 - the so-called IDCM feature uses an eligibility criteria prior
   to permitting an encoder mode decision.

 - the feature may be controlled via the command line / config file
   using the --inferredDirectCodingMode=1|0 option

This commit provides an implementation of m42238 to code occupancy:

 - a node's child occupancy is contextualised based on the adjacent
   neighbours of the parent node.

 - when an isolated node is occupied by only a single child, the
   sub-node address [0,7] is directly coded instead of the 8-bit
   occupancy word.

 - collection of neighbour occupancy is performed using a scatter/
   gather operation upon the insertion of each child node into the
   fifo queue.

 - the feature may be controlled via the command line / config file
   using the --neighbourContextualisation=1|0 option

Toshiyasu Sugio authored Apr 24, 2018 and David Flynn committed May 10, 2018

In cases where all point positions are unique (either due to input or
due to mergeDuplicatePoints = 1), there is no need to signal the number
of points in each leaf node (since it is always 1).

This commit adds a header flag to indicate this case, and modifies the
encoder and decoder to omit the point count signalling when
mergeDuplicatePoints = 1.

This change avoids enqueueing an additional node for every point in the
input for the sole purpose of encoding the position count in a
subsequent iteration.

At the penultimate level of the geometry octree, a node's children
(which cannot be further split) are immediately encoded or decoded.

This commit eliminates the temporary fifo used in partitioning each node
into child nodes.  Rather than using independent sorting passes to apply
a binary partition to each of the x, y, and z axis, a counting sort
permits all eight output positions to be calculated with a single pass,
and a further pass to reorder the elements.

Use of an std::deque can involve a large number of malloc operations
as the internal sections of the queue are repeatedly allocated and
released during the geometry coding loop.

This commit provides a ring buffer implementation that uses the
worst-case octree node count (ie, ~number of input points) as
a fixed-size buffer.

This commit replaces the use of size_t (the index type of the input
point list) with an explicit uint32_t type on the basis that it is
not currently feasible to consider a point cloud with more than
2**32 points.

uint32_t is used to follow the STL convention of a vector's size_type
being unsigned.

This optimisation reduces the size of the PCCOctree3Node structure
by replacing the per-node bounding box with the node's position (which
is the same as the bounding box minimums).  The contribution of the
bounding box maximums can be deduced from the initial bounding box
and the current tree depth.

Counters are used to track the transition from one depth to the next
during breadth-first traversal.

This commit fixes a bug where coding certain point clouds may
result in points being moved at the edge of the bounding box.
One such manifestation is the appearance of duplicate points in
a point cloud originally with unique point positions.

The maximum should be calculated as the lowest power of two minus
one that contains the bounding box.