Yiting Shao authored Dec 07, 2018 and David Flynn committed Feb 06, 2019

This partitioning method (--partitionMethod=2) finds the longest edge of
the point cloud and divides it into --partitionNumUniformGeom=n slices
along the longest edge.

If n = 0, the ratio of longest edge to shortest edge determines
the number of slices.

This commit provides a basic frame work for partitioning a frame into
multiple slices, with the continued assumption of single-frame
sequences.

The decoder is modified to independently decode each slice and
accumulate decoded points in a buffer for output.

The encoder is updated to support partitioning the input point cloud
into slices and to independently code each slice.  Points from
reconstructed slices are accumulated and output at the end of the
frame period.

The partitioning process (partitioning methods are defined in
partitioning.cpp) proceeds as follows:

 - quantise the input point cloud without duplicate point removal
   or reordering points.

 - apply the partitioning function to produce a list of tiles
   and slices, each slice having an origin, id, and list of point
   indexes that identify points in the input point cloud.

 - producing a source point cloud for each partition as a subset
   of the input point cloud.

 - compressing each partition (slice) as normal by quantising
   the partitioned input.  Recolouring is necessarily performed
   against the partitioned input since the recolouring method
   cannot correctly handle recolouring a partition from a complete
   point cloud.

NB: this commit does not provide any partitioning methods.

This commit provides support for each slice to have an origin
specified relative to the sequence bounding box.

Reconstructed points are offset by the slice origin prior to inverse
scaling and translation.

With slice/tile-based coding, it is necessary to build an output
point cloud from independent slice/tiles.  An append operation allows
the points of two point clouds to be concatenated.

Vladyslav Zakharchenko authored Dec 14, 2018 and David Flynn committed Feb 06, 2019

This uses a binary-tree to generate the levels-of-details for the
lifting/predicting transforms as an alternative to Euclidean distance
thresholding methods.

Jungsun Kim authored Jan 30, 2019 and David Flynn committed Feb 05, 2019

This simplified attribute prediction scheme avoids computing the
LoD structure in order to reduce the computational complexity of
both the encoder and the decoder.  It is activated by setting
levelOfDetailCount=0.

NB: this is only configured for lossless/near-lossless predictive
attribute coding of cat3 sequences.

Danillo Bracco Graziosi authored Dec 12, 2018 and David Flynn committed Feb 05, 2019

This replaces the previous floating point transform implementation with
a fixed-point alternative with essentially identical compression
performance.

Ohji Nakagami authored Feb 04, 2019 and David Flynn committed Feb 05, 2019

This commit provides an integer approximation of atan2().

Ohji Nakagami authored Nov 26, 2018 and David Flynn committed Feb 05, 2019

This replaces the previous floating point version.  The number of
fractional bits is set to 8, allowing 24bit (unsigned) geometry
to be represented in 32bit calculations with negligible effects
on the reconstruction.

 - calculates the sequence bounding box at the start of encoding
   (in original coordinates as a decoder may reconstruct it)
 - signals bounding box in SPS
 - disables signalling of geometry box origin (always 0)

The configuration parameters seq_bounding_box_xyz0 and
seq_bounding_box_whd allow overriding the automatic derivation of
frame (sequence) bounding box, forcing the depth of the geometry
tree.

NB: seq_bounding_box_xyz0 and seq_bounding_box_whd are specified
    using unscaled co-ordinates.

NB: this patch does not signal these parameters in the SPS.

Define po-lite option traits to handle PCCVector3<T> as a container
to permit using PCCVector3<T> types as command / config parameters.

 - signal geom_slice_id / attr_geom_slice_id in geometry / attribute
   header.

 - attr_geom_slice_id set to current geom_slice_id to maintain
   relationship between attribute and geometry.

 - increment geom_slice_id after each round of encoding.

 - decoder checks that the attribute payload's attr_geom_slice_id
   corresponds to the most recently decoded geometry payload.

Each occupancy bit is contextualised based on the number of occupied
(directly) adjacent child nodes in the (x-1), (y-1), and (z-1)
neighbour nodes used to generate the neighbour pattern.

Three contextualisation states are added:
 - no adjacent neighbours
 - one adjacent neigbour (=adjacencyGt0)
 - two adjacent neighoburs (=adjacencyGt1)

The neighbour pattern for a node indicates occupied neighbouring nodes.

This commit updates the definition of occupancy for the (x-1), (y-1),
and (z-1) neighbours to be occupied if the neighbour's adjoining
children (ie sharing a face with the current node) are occupied.

In the 1-D figure below, at a depth d, the node 'X' has two occupied
neighbours 'R' and 'L'.  When determining the neighbour pattern of X,
the (x-1) state of the neighbour's adjoining child ('r') is used
instead.  For completeness, nodes marked with '?' are not causally
available.

  d+0:   R|X|L
  d+1:  Rr|??|??

Replaces the existing map update tables (generated from (L*p+s)/(L+1),
memory L=10) with an adaptive memory (L=max(5, 1/p, 1/(1-p)) clipped
to 200) represented as delta transitions and quantised to 16 entries.

This commit provides trisoup_node_size_log2 as a replacement for
the following HLS syntax variables:
 - geom_codec_type (redundant wrt. triangle node size derivation)
 - trisoup_depth (redundant wrt. bounding box)
 - trisoup_triangle_level (semantics simplified)

The following config options have been removed:
 - geometryCodec (redundant wrt. trisoup_node_size_log2)
 - triSoupLevel  (replaced by trisoup_node_size_log2)
 - triSoupDepth  (unnecessary)
 - triSoupIntToOrigScale (redundant wrt. positionQuantizationScale)

Min Ku Lee authored Jan 30, 2019 and David Flynn committed Feb 05, 2019

Using the Rec.709 colour matrix with full range RGB signals, it is
necessary to clip the output to avoid exceeding the attribute bitdepth.

A leaf node by definition contains at least one point.  Therefore the
single point flag indicates if there are one or two points in the leaf.
The remainder should be signalled as numpoints - 2 as per the working
draft text.

Khaled Mammou authored Sep 22, 2018 and David Flynn committed Nov 01, 2018

Replace the k-d tree based nearest neighbour search with an approximate
search based upon the points in Morton order.

Toshiyasu Sugio authored Sep 20, 2018 and David Flynn committed Nov 01, 2018

This commit provides an ability to vary the number of direct neighbours
used in the predicting transform.  Mode 0 corresponds to the existing
averaging, and mode n selects the n-th nearest neighbour for direct
prediction.

The provided code has been reworked to:
 - rewrite unary coding to be a little more obvious
 - tidy the mode decision path
 - add configuration for max_num_direct_predictors (renamed from
   MaxNumPredCand)
 - now counts the number of single predictors, rather than prediction
   modes.

This commit provides an implementation of the entropy coding interface
using the dirac (schroedinger) arithmetic codec.

In order to handle any remaining m-ary symbols, a naïve unary
binarisation is employed.

This commit adds an arithmetic codec interface class that allows a
compile time choice of arithmetic codec implementation.  Context types
are renamed to support compile time selection, and existing support
functions that were added to the third-party arithmetic codec are
moved to the wrapper.

This commit provides a method to predict the child occupancy bits of a
node based on the node's 26 neighbours.  The prediction is used to
contextualise coding of each occupancy bit.

This tool requires the use of the occupancyAtlas for neighbour lookup.
NB: a restriction in the current implementation requires that the
atlas size is at most 8³.

intra_pred_max_node_size_log2: 6

This commit provides the ability to contextualise occupancy bits using
a prediction.  Three context sets cover the not predicted, predicted
unoccupied and predicted occupied cases.

chi wang authored Oct 09, 2018 and David Flynn committed Nov 01, 2018

When restricted neighbour contextualisation is used, this modification
introduces an alternative neighbour reduction table.

NB: the integrated table has been updated to remove the need to add
additional remapping tables.  This is a non-normative change.

Bitwise geometry occupancy coding uses a mapping table to select entropy
contexts.  This commit employs a dynamic mapping which is updated after
each coding operation, replacing the previous static mapping tables.

NB: the proposed version used a context with a halving period
(max_count) of 64 symbols.  However, this conflicts with another
adoption (512 symbols), and a wholesale replacement of the arithmetic
codec and context model.  To resolve the conflict, the existing halving
period (128) is used.

Khaled Mammou authored Sep 18, 2018 and David Flynn committed Nov 01, 2018

This commit replaces the m-ary coding of transform coefficient values
with a binary coding using the dual-lut coder.

Reworked to include a bug fix where the maximum attribute symbol value
used by the dual lut coder is 64 rather than 255

Khaled Mammou authored Sep 19, 2018 and David Flynn committed Nov 01, 2018

This commit allows bytewise encoding of occupancy symbols using the
dual-lut coder.

Khaled Mammou authored Sep 19, 2018 and David Flynn committed Nov 01, 2018

This commit provides an m-ary entropy coder based on a fixed-size
dictionary with periodic updates, a cache of recently used symbols
(updated using an LRU eviction policy), and a fallback direct binary
coding of any unhandled symbols.

NB: the proposed version used a context with a halving period
(max_count) of 64 symbols.  However, this conflicts with another
adoption (512 symbols), and a wholesale replacement of the arithmetic
codec and context model.  To resolve the conflict, the existing halving
period (128) is used.

This commit provides an encapsulated interface to the fast update option
of the o3dgc context models with a statistic halving period of 128
symbols (max_count).

This commit prepares for the introduction of selectable bitwise
and bytewise occupancy coding methods.

Ohji Nakagami authored Aug 07, 2018 and David Flynn committed Nov 01, 2018

This commit uses a direct intersection check between the trisoup
triangles and point cloud unit vectors.  This avoids the need to
upsample the triangles with dense sampling and then performing
voxelisation.

Philip A. Chou authored Jul 19, 2018 and David Flynn committed Nov 01, 2018

This commit integrates a c++ trisoup codec, replacing the previous matlab
implementation.

The provided code has been reworked to avoid duplicating code, dead code,
and operate with the current HLS.

Ohji Nakagami authored Jul 26, 2018 and David Flynn committed Oct 31, 2018

When trisoup was completely independent, trisoup handled lossless
geometry coding by using a leaf size of 1 and no triangles.  This
commit replaces that mode with the better performing octree geometry
coder.