geom/m49231: add support for non-cubic nodes with implicit qtbt

This adoption permits coding geometry with non-cubic bounding boxes.
Since the depth of the tree remains constant for cubic and non-cubic
bounding boxes with identical largest dimensions, quad-tree and
binary-tree partitions are introduced to avoid coding 'fractional'
positions.

The following configuration options control the placement of non-octree
partitions within the coding tree:

 --max_num_implicit_qtbt_before_ot
 --min_implicit_qtbt_size_log2

cfg/octree-liftt-ctc-lossless-geom-lossy-attrs.yaml

@@ -16,6 +16,13 @@ categories:
      - intra_pred_max_node_size_log2: 6
      - positionQuantizationScale: 1
 
+     # default qtbt with cat3-frame exception
+     - max_num_implicit_qtbt_before_ot: 4
+     - min_implicit_qtbt_size_log2: 0
+     -
+       - !conditional '"${group}" eq "cat3-frame"'
+       - max_num_implicit_qtbt_before_ot: 0
+
      ####
      # attribute coding (common options -- relies on option ordering)
      #  - use lifting transform for lossy conditions

cfg/octree-liftt-ctc-lossy-geom-lossy-attrs.yaml

@@ -21,6 +21,13 @@ categories:
          r05: '$eval{ $rp = 1; $gp = ${src-geometry-precision}; $p_min = max(gp - 9, 7); $start = min(1, $gp - ($p_min + 6)); $step = max(1, (min($gp - 1, $p_min + 7) - $p_min) / 5); $y = $start + round($rp * $step); $div = 1 << (abs($y) + 1); ((1 - 2*signbit($y)) % $div) / $div }'
          r06: '$eval{ $rp = 0; $gp = ${src-geometry-precision}; $p_min = max(gp - 9, 7); $start = min(1, $gp - ($p_min + 6)); $step = max(1, (min($gp - 1, $p_min + 7) - $p_min) / 5); $y = $start + round($rp * $step); $div = 1 << (abs($y) + 1); ((1 - 2*signbit($y)) % $div) / $div }'
 
+     # default qtbt with cat3-frame exception
+     - max_num_implicit_qtbt_before_ot: 4
+     - min_implicit_qtbt_size_log2: 0
+     -
+       - !conditional '"${group}" eq "cat3-frame"'
+       - max_num_implicit_qtbt_before_ot: 0
+
      ####
      # attribute coding (common options -- relies on option ordering)
      #  - skip attribute coding if dist2 parameters are not defined

cfg/octree-predt-ctc-lossless-geom-lossless-attrs.yaml

@@ -17,6 +17,13 @@ categories:
      - neighbourAvailBoundaryLog2: 8
      - intra_pred_max_node_size_log2: 6
 
+     # default qtbt with cat3-frame exception
+     - max_num_implicit_qtbt_before_ot: 4
+     - min_implicit_qtbt_size_log2: 0
+     -
+       - !conditional '"${group}" eq "cat3-frame"'
+       - max_num_implicit_qtbt_before_ot: 0
+
      ####
      # attribute coding (common options -- relies on option ordering)
      #  - skip attribute coding if dist2 parameters are not defined

cfg/octree-predt-ctc-lossless-geom-nearlossless-attrs.yaml

@@ -17,6 +17,13 @@ categories:
      - neighbourAvailBoundaryLog2: 8
      - intra_pred_max_node_size_log2: 6
 
+     # default qtbt with cat3-frame exception
+     - max_num_implicit_qtbt_before_ot: 4
+     - min_implicit_qtbt_size_log2: 0
+     -
+       - !conditional '"${group}" eq "cat3-frame"'
+       - max_num_implicit_qtbt_before_ot: 0
+
      ####
      # attribute coding (common options -- relies on option ordering)
      #  - code directly GBR (no need to perform colourspace conversion)

cfg/octree-raht-ctc-lossless-geom-lossy-attrs.yaml

@@ -16,6 +16,13 @@ categories:
      - intra_pred_max_node_size_log2: 6
      - positionQuantizationScale: 1
 
+     # default qtbt with cat3-frame exception
+     - max_num_implicit_qtbt_before_ot: 4
+     - min_implicit_qtbt_size_log2: 0
+     -
+       - !conditional '"${group}" eq "cat3-frame"'
+       - max_num_implicit_qtbt_before_ot: 0
+
      ####
      # attribute coding (common options -- relies on option ordering)
      #  - uses raht transform

cfg/octree-raht-ctc-lossy-geom-lossy-attrs.yaml

@@ -21,6 +21,13 @@ categories:
          r05: '$eval{ $rp = 1; $gp = ${src-geometry-precision}; $p_min = max(gp - 9, 7); $start = min(1, $gp - ($p_min + 6)); $step = max(1, (min($gp - 1, $p_min + 7) - $p_min) / 5); $y = $start + round($rp * $step); $div = 1 << (abs($y) + 1); ((1 - 2*signbit($y)) % $div) / $div }'
          r06: '$eval{ $rp = 0; $gp = ${src-geometry-precision}; $p_min = max(gp - 9, 7); $start = min(1, $gp - ($p_min + 6)); $step = max(1, (min($gp - 1, $p_min + 7) - $p_min) / 5); $y = $start + round($rp * $step); $div = 1 << (abs($y) + 1); ((1 - 2*signbit($y)) % $div) / $div }'
 
+     # default qtbt with cat3-frame exception
+     - max_num_implicit_qtbt_before_ot: 4
+     - min_implicit_qtbt_size_log2: 0
+     -
+       - !conditional '"${group}" eq "cat3-frame"'
+       - max_num_implicit_qtbt_before_ot: 0
+
      ####
      # attribute coding (common options -- relies on option ordering)
      #  - use raht

doc/README.options.md

@@ -235,6 +235,18 @@ A non-normative encoder process determines the QP offset based upon
 the local density of the octree.  A value of -1 disables signalling
 of per-node QP offsets.
 
+### `implicitQtBtEnabled=0|1`
+Enables non-cubic geometry tree coding.  When enabled, the geometry
+tree may have a cuboid bounding box.  Internal tree nodes are
+implicitly determined to be octree, quadtree or binary tree nodes.
+
+### `max_num_implicit_qtbt_before_ot=INT-VALUE`
+Limits the maximal number of implicit quadtree and binary tree
+partitions used before the first octree partition.
+
+### `min_implicit_qtbt_size_log2=INT-VALUE`
+Specifies the minimum size of implicit quadtree and binary tree partitions
+
 ### `--bitwiseOccupancyCoding=0|1`
 In octree geometry coding, there are both byte-wise and bit-wise tools to
 encode the occupancy data.  This option selects between the two methods.

tmc3/OctreeNeighMap.cpp

@@ -44,6 +44,7 @@ namespace pcc {
 void
 updateGeometryOccupancyAtlas(
   const Vec3<uint32_t>& currentPosition,
+  const int atlasShift,
   const pcc::ringbuf<PCCOctree3Node>& fifo,
   const pcc::ringbuf<PCCOctree3Node>::iterator& fifoCurrLvlEnd,
   MortonMap3D* occupancyAtlas,
@@ -51,6 +52,10 @@ updateGeometryOccupancyAtlas(
 {
   const uint32_t mask = (1 << occupancyAtlas->cubeSizeLog2()) - 1;
   const int shift = occupancyAtlas->cubeSizeLog2();
+  const int shiftX = (atlasShift & 4 ? 1 : 0);
+  const int shiftY = (atlasShift & 2 ? 1 : 0);
+  const int shiftZ = (atlasShift & 1 ? 1 : 0);
+
   const auto currentOrigin = currentPosition >> shift;
 
   // only refresh the atlas if the current position lies outside the
@@ -65,10 +70,9 @@ updateGeometryOccupancyAtlas(
   for (auto it = fifo.begin(); it != fifoCurrLvlEnd; ++it) {
     if (currentOrigin != it->pos >> shift)
       break;
-
-    const uint32_t x = it->pos[0] & mask;
-    const uint32_t y = it->pos[1] & mask;
-    const uint32_t z = it->pos[2] & mask;
+    const uint32_t x = (it->pos[0] & mask) >> shiftX;
+    const uint32_t y = (it->pos[1] & mask) >> shiftY;
+    const uint32_t z = (it->pos[2] & mask) >> shiftZ;
     occupancyAtlas->setByte(x, y, z, it->siblingOccupancy);
   }
 }
@@ -138,6 +142,7 @@ GeometryNeighPattern
 makeGeometryNeighPattern(
   bool adjacent_child_contextualization_enabled_flag,
   const Vec3<uint32_t>& position,
+  const int atlasShift,
   const MortonMap3D& occupancyAtlas)
 {
   const int mask = occupancyAtlas.cubeSize() - 1;
@@ -146,6 +151,7 @@ makeGeometryNeighPattern(
   const int32_t y = position[1] & mask;
   const int32_t z = position[2] & mask;
   uint8_t neighPattern;
+  if (atlasShift == 0) {
     if (
       x > 0 && x < cubeSizeMinusOne && y > 0 && y < cubeSizeMinusOne && z > 0
       && z < cubeSizeMinusOne) {
@@ -163,6 +169,40 @@ makeGeometryNeighPattern(
       neighPattern |= occupancyAtlas.getWithCheck(x, y, z - 1) << 4;
       neighPattern |= occupancyAtlas.getWithCheck(x, y, z + 1) << 5;
     }
+  } else {
+    const int shiftX = (atlasShift & 4 ? 1 : 0);
+    const int shiftY = (atlasShift & 2 ? 1 : 0);
+    const int shiftZ = (atlasShift & 1 ? 1 : 0);
+
+    if (
+      x > 0 && x < cubeSizeMinusOne && y > 0 && y < cubeSizeMinusOne && z > 0
+      && z < cubeSizeMinusOne) {
+      neighPattern = occupancyAtlas.get(x + 1, y, z, shiftX, shiftY, shiftZ);
+      neighPattern |= occupancyAtlas.get(x - 1, y, z, shiftX, shiftY, shiftZ)
+        << 1;
+      neighPattern |= occupancyAtlas.get(x, y - 1, z, shiftX, shiftY, shiftZ)
+        << 2;
+      neighPattern |= occupancyAtlas.get(x, y + 1, z, shiftX, shiftY, shiftZ)
+        << 3;
+      neighPattern |= occupancyAtlas.get(x, y, z - 1, shiftX, shiftY, shiftZ)
+        << 4;
+      neighPattern |= occupancyAtlas.get(x, y, z + 1, shiftX, shiftY, shiftZ)
+        << 5;
+    } else {
+      neighPattern =
+        occupancyAtlas.getWithCheck(x + 1, y, z, shiftX, shiftY, shiftZ);
+      neighPattern |=
+        occupancyAtlas.getWithCheck(x - 1, y, z, shiftX, shiftY, shiftZ) << 1;
+      neighPattern |=
+        occupancyAtlas.getWithCheck(x, y - 1, z, shiftX, shiftY, shiftZ) << 2;
+      neighPattern |=
+        occupancyAtlas.getWithCheck(x, y + 1, z, shiftX, shiftY, shiftZ) << 3;
+      neighPattern |=
+        occupancyAtlas.getWithCheck(x, y, z - 1, shiftX, shiftY, shiftZ) << 4;
+      neighPattern |=
+        occupancyAtlas.getWithCheck(x, y, z + 1, shiftX, shiftY, shiftZ) << 5;
+    }
+  }
 
   // Above, the neighbour pattern corresponds directly to the six same
   // sized neighbours of the given node.

tmc3/OctreeNeighMap.h

@@ -62,7 +62,7 @@ public:
     const uint32_t halfCubeSizeLog2 = cubeSizeLog2 ? cubeSizeLog2 - 1 : 0;
     _cubeSizeLog2 = cubeSizeLog2;
     _cubeSize = 1 << cubeSizeLog2;
-    _bufferSizeInBytes = 1 << (3 * halfCubeSizeLog2);
+    _bufferSizeInBytes = 1 << (3 * cubeSizeLog2);
     _buffer.reset(new uint8_t[_bufferSizeInBytes]);
     _childOccupancy.reset(new uint8_t[_bufferSizeInBytes << 3]);
     _updates.reserve(1 << 16);
@@ -98,12 +98,44 @@ public:
     }
   }
 
+  uint32_t get(
+    const int32_t x,
+    const int32_t y,
+    const int32_t z,
+    const int shiftX,
+    const int shiftY,
+    const int shiftZ) const
+  {
+    assert(
+      x >= 0 && y >= 0 && z >= 0 && x < _cubeSize && y < _cubeSize
+      && z < _cubeSize);
+    return (_buffer[getByteIndex(x >> shiftX, y >> shiftY, z >> shiftZ)]
+            >> getBitIndex(shiftX ? x : 0, shiftY ? y : 0, shiftZ ? z : 0))
+      & 1;
+  }
+
+  uint32_t getWithCheck(
+    const int32_t x,
+    const int32_t y,
+    const int32_t z,
+    const int shiftX,
+    const int shiftY,
+    const int shiftZ) const
+  {
+    if (
+      x < 0 || x >= _cubeSize || y < 0 || y >= _cubeSize || z < 0
+      || z >= _cubeSize) {
+      return false;
+    }
+    return get(x, y, z, shiftX, shiftY, shiftZ);
+  }
+
   uint32_t get(const int32_t x, const int32_t y, const int32_t z) const
   {
     assert(
       x >= 0 && y >= 0 && z >= 0 && x < _cubeSize && y < _cubeSize
       && z < _cubeSize);
-    return (_buffer[getByteIndex(x, y, z)] >> getBitIndex(x, y, z)) & 1;
+    return _buffer[getByteIndex(x, y, z)] & 1;
   }
 
   uint32_t
@@ -119,13 +151,12 @@ public:
 
   void setChildOcc(int32_t x, int32_t y, int32_t z, uint8_t childOccupancy)
   {
-    _childOccupancy[getByteIndex(x << 1, y << 1, z << 1)] = childOccupancy;
+    _childOccupancy[getByteIndex(x, y, z)] = childOccupancy;
   }
 
   uint8_t getChildOcc(int32_t x, int32_t y, int32_t z) const
   {
-    uint8_t childOccupancy =
-      _childOccupancy[getByteIndex(x << 1, y << 1, z << 1)];
+    uint8_t childOccupancy = _childOccupancy[getByteIndex(x, y, z)];
     return childOccupancy;
   }
 
@@ -138,8 +169,7 @@ private:
   uint32_t
   getByteIndex(const int32_t x, const int32_t y, const int32_t z) const
   {
-    return kMortonCode256X[x >> 1] | kMortonCode256Y[y >> 1]
-      | kMortonCode256Z[z >> 1];
+    return kMortonCode256X[x] | kMortonCode256Y[y] | kMortonCode256Z[z];
   }
 
   int _cubeSize = 0;
@@ -177,12 +207,14 @@ struct GeometryNeighPattern {
 GeometryNeighPattern makeGeometryNeighPattern(
   bool adjacent_child_contextualization_enabled_flag,
   const Vec3<uint32_t>& currentPosition,
+  const int atlasShift,
   const MortonMap3D& occupancyAtlas);
 
 // populate (if necessary) the occupancy atlas with occupancy information
 // from @fifo.
 void updateGeometryOccupancyAtlas(
   const Vec3<uint32_t>& position,
+  const int atlasShift,
   const ringbuf<PCCOctree3Node>& fifo,
   const ringbuf<PCCOctree3Node>::iterator& fifoCurrLvlEnd,
   MortonMap3D* occupancyAtlas,

tmc3/PCCMath.h

@@ -46,6 +46,8 @@
 #include "PCCMisc.h"
 #include "tables.h"
 
+#include <algorithm>
+
 namespace pcc {
 /// Vector dim 3
 template<typename T>

tmc3/TMC3.cpp

@@ -506,6 +506,18 @@ ParseParameters(int argc, char* argv[], Parameters& params)
 
   (po::Section("Geometry"))
 
+  ("implicitQtBtEnabled",
+    params.encoder.gps.implicit_qtbt_enabled_flag, true,
+    "Enables non-cubic geometry bounding box")
+
+  ("max_num_implicit_qtbt_before_ot",
+    params.encoder.gps.max_num_implicit_qtbt_before_ot, 4,
+    "Max number of implicit qtbt before ot")
+
+  ("min_implicit_qtbt_size_log2",
+    params.encoder.gps.min_implicit_qtbt_size_log2, 0,
+    "Minimum size of implicit qtbt")
+
   // tools
   ("bitwiseOccupancyCoding",
     params.encoder.gps.bitwise_occupancy_coding_flag, true,

tmc3/encoder.cpp

@@ -451,13 +451,6 @@ void
 PCCTMC3Encoder3::encodeGeometryBrick(
   const EncoderParams* params, PayloadBuffer* buf)
 {
-  // todo(df): confirm minimum of 1 isn't needed
-  int32_t maxBB =
-    std::max({1, _sliceBoxWhd[0], _sliceBoxWhd[1], _sliceBoxWhd[2]});
-
-  // the current node dimension (log2) encompasing maxBB
-  int nodeSizeLog2 = ceillog2(maxBB + 1);
-
   GeometryBrickHeader gbh;
   gbh.geom_geom_parameter_set_id = _gps->gps_geom_parameter_set_id;
   gbh.geom_slice_id = _sliceId;
@@ -465,7 +458,19 @@ PCCTMC3Encoder3::encodeGeometryBrick(
   gbh.frame_idx = _frameCounter & ((1 << _sps->log2_max_frame_idx) - 1);
   gbh.geomBoxOrigin = _sliceOrigin;
   gbh.geom_box_log2_scale = 0;
-  gbh.geom_max_node_size_log2 = nodeSizeLog2;
+
+  if (!_gps->implicit_qtbt_enabled_flag) {
+    // todo(df): confirm minimum of 1 isn't needed
+    int32_t maxBB =
+      std::max({1, _sliceBoxWhd[0], _sliceBoxWhd[1], _sliceBoxWhd[2]});
+    gbh.geom_max_node_size_log2 = ceillog2(maxBB + 1);
+  } else {
+    // different node dimension for xyz, for the purpose of implicit qtbt
+    gbh.geom_max_node_size_log2_xyz[0] = ceillog2(_sliceBoxWhd[0] + 1);
+    gbh.geom_max_node_size_log2_xyz[1] = ceillog2(_sliceBoxWhd[1] + 1);
+    gbh.geom_max_node_size_log2_xyz[2] = ceillog2(_sliceBoxWhd[2] + 1);
+  }
+
   gbh.geom_num_points = int(pointCloud.getPointCount());
   gbh.geom_slice_qp_offset = params->gbh.geom_slice_qp_offset;
   gbh.geom_octree_qp_offset_enabled_flag =

tmc3/geometry_intra_pred.cpp

@@ -63,6 +63,7 @@ void
 predictGeometryOccupancyIntra(
   const MortonMap3D& occupancyAtlas,
   Vec3<uint32_t> pos,
+  const int atlasShift,
   int* occupancyIsPredicted,
   int* occupancyPrediction)
 {
@@ -75,7 +76,7 @@ predictGeometryOccupancyIntra(
   int numOccupied = 0;
 
   const int8_t* p = &LUT_dist[0][0];
-
+  if (atlasShift == 0) {
     for (int dx = -1; dx <= 1; dx++) {
       for (int dy = -1; dy <= 1; dy++) {
         for (int dz = -1; dz <= 1; dz++) {
@@ -85,6 +86,34 @@ predictGeometryOccupancyIntra(
           // todo(df): remove unnecessary checks
           bool occupied = occupancyAtlas.getWithCheck(x + dx, y + dy, z + dz);
 
+          if (occupied) {
+            for (int i = 0; i < 8; i++)
+              score[i] += LUT1[p[i]];
+            numOccupied++;
+          } else {
+            for (int i = 0; i < 8; i++)
+              score[i] += LUT0[p[i]];
+          }
+          // next pattern
+          p += 8;
+        }
+      }
+    }
+  } else {
+    const int shiftX = (atlasShift & 4 ? 1 : 0);
+    const int shiftY = (atlasShift & 2 ? 1 : 0);
+    const int shiftZ = (atlasShift & 1 ? 1 : 0);
+
+    for (int dx = -1; dx <= 1; dx++) {
+      for (int dy = -1; dy <= 1; dy++) {
+        for (int dz = -1; dz <= 1; dz++) {
+          if (dz == 0 && dy == 0 && dx == 0)
+            continue;
+
+          // todo(df): remove unnecessary checks
+          bool occupied = occupancyAtlas.getWithCheck(
+            x + dx, y + dy, z + dz, shiftX, shiftY, shiftZ);
+
           if (occupied) {
             for (int i = 0; i < 8; i++)
               score[i] += LUT1[p[i]];
@@ -99,6 +128,7 @@ predictGeometryOccupancyIntra(
         }
       }
     }
+  }
 
   if (numOccupied <= 8) {
     *occupancyIsPredicted = 0;

tmc3/geometry_intra_pred.h

@@ -47,6 +47,7 @@ namespace pcc {
 void predictGeometryOccupancyIntra(
   const MortonMap3D& occupancyAtlas,
   Vec3<uint32_t> pos,
+  const int atlasShift,
   int* occupacyIsPredIntra,
   int* occupacyPredIntra);
 

tmc3/geometry_octree.h

@@ -114,6 +114,18 @@ void decodeGeometryOctree(
   EntropyDecoder* arithmeticDecoder,
   pcc::ringbuf<PCCOctree3Node>* nodesRemaining);
 
+//---------------------------------------------------------------------------
+// Determine if a node is a leaf node based on size.
+// A node with all dimension = 0 is a leaf node.
+// NB: some dimensions may be less than zero if coding of that dimension
+// has already terminated.
+
+inline bool
+isLeafNode(const Vec3<int>& sizeLog2)