Commit 8f3fd506 authored by David Flynn's avatar David Flynn
Browse files

geom: fix interaction of qtbt and octree geometry quantisation

The geometry quantisation assumed that octree nodes were cubes.  The
addition of qt/bt partitioning modes breaks this premise.  This commit
addresses the following:

 - Only the to-be-coded bits, which may differ for each component, are
   quantised, rather than quantising all bits according to the tree
   depth (which would result in quantising already coded bits).

 - The occupancy mask (determining the non-coded occupancy bits) must
   be recomputed for each node.  Since the effect of quantisation is
   to prune back each sub-tree from the leaves, some nodes that would
   have had qt/bt partitions will not be coded, leaving an unequal
   reconstruction.  The solution is to terminate the shorter leaves
   earlier and to recompute the partition.
parent 667f9209
......@@ -475,7 +475,7 @@ GeometryOctreeDecoder::decodeDirectPosition(
// Helper to inverse quantise positions
Vec3<uint32_t>
invQuantPosition(int qp, int quantBitMask, const Vec3<uint32_t>& pos)
invQuantPosition(int qp, Vec3<uint32_t> quantMasks, const Vec3<uint32_t>& pos)
{
// pos represents the position within the coded tree as follows:
// |pppppqqqqqq|00
......@@ -489,7 +489,7 @@ invQuantPosition(int qp, int quantBitMask, const Vec3<uint32_t>& pos)
int shiftBits = (qp - 4) / 6;
Vec3<uint32_t> recon;
for (int k = 0; k < 3; k++) {
int posQuant = pos[k] & (quantBitMask >> shiftBits);
int posQuant = pos[k] & (quantMasks[k] >> shiftBits);
recon[k] = (pos[k] ^ posQuant) << shiftBits;
recon[k] |= quantizer.scale(posQuant);
}
......@@ -570,15 +570,12 @@ decodeGeometryOctree(
int sliceQp = gps.geom_base_qp + gbh.geom_slice_qp_offset;
int numLvlsUntilQpOffset = -1;
int posQuantBits = 0;
Vec3<uint32_t> posQuantBitMasks = 0xffffffff;
if (gbh.geom_octree_qp_offset_enabled_flag)
numLvlsUntilQpOffset = gbh.geom_octree_qp_offset_depth;
else if (gps.geom_scaling_enabled_flag) {
else if (gps.geom_scaling_enabled_flag)
node00.qp = sliceQp;
// determine the mask of LSBs used in quantisation
posQuantBits = (1 << (nodeMaxDimLog2 - numLvlsUntilQpOffset)) - 1;
}
for (; !fifo.empty(); fifo.pop_front()) {
if (fifo.begin() == fifoCurrLvlEnd) {
......@@ -596,7 +593,7 @@ decodeGeometryOctree(
if (maxNumImplicitQtbtBeforeOt)
maxNumImplicitQtbtBeforeOt--;
// if all dimensions have same size, then performing octree for remaining nodes
// if all dimensions have same size, then use octree for remaining nodes
if (
nodeSizeLog2[0] == nodeSizeLog2[1]
&& nodeSizeLog2[1] == nodeSizeLog2[2])
......@@ -606,8 +603,6 @@ decodeGeometryOctree(
childSizeLog2 = implicitQtBtDecision(
nodeSizeLog2, maxNumImplicitQtbtBeforeOt, minSizeImplicitQtbt);
occupancySkip = nonSplitQtBtAxes(nodeSizeLog2, childSizeLog2);
nodeMaxDimLog2--;
numNodesNextLvl = 0;
occupancyAtlasOrigin = 0xffffffff;
......@@ -622,7 +617,12 @@ decodeGeometryOctree(
if (nodeMaxDimLog2 == minNodeSizeLog2)
break;
// record the node size when quantisation is signalled -- all subsequnt
// coded occupancy bits are quantised
numLvlsUntilQpOffset--;
if (!numLvlsUntilQpOffset)
for (int k = 0; k < 3; k++)
posQuantBitMasks[k] = (1 << nodeSizeLog2[k]) - 1;
}
PCCOctree3Node& node0 = fifo.front();
......@@ -635,6 +635,16 @@ decodeGeometryOctree(
auto effectiveNodeSizeLog2 = nodeSizeLog2 - shiftBits;
auto effectiveChildSizeLog2 = childSizeLog2 - shiftBits;
// todo(??): the following needs to be reviewed, it is added to make
// quantisation work with qtbt.
Vec3<int> actualNodeSizeLog2, actualChildSizeLog2;
for (int k = 0; k < 3; k++) {
actualNodeSizeLog2[k] = std::max(nodeSizeLog2[k], shiftBits);
actualChildSizeLog2[k] = std::max(childSizeLog2[k], shiftBits);
}
// todo(??): atlasShift may be wrong too
occupancySkip = nonSplitQtBtAxes(actualNodeSizeLog2, actualChildSizeLog2);
int occupancyAdjacencyGt0 = 0;
int occupancyAdjacencyGt1 = 0;
int occupancyAdjacencyUnocc = 0;
......@@ -710,7 +720,7 @@ decodeGeometryOctree(
(node0.pos[1] << !(occupancySkip & 2)) + y,
(node0.pos[2] << !(occupancySkip & 1)) + z};
pos = invQuantPosition(node0.qp, posQuantBits, pos);
pos = invQuantPosition(node0.qp, posQuantBitMasks, pos);
const Vec3<double> point(pos[0], pos[1], pos[2]);
for (int i = 0; i < numPoints; ++i)
......@@ -747,7 +757,7 @@ decodeGeometryOctree(
point[k] += child.pos[k] << shift;
}
point = invQuantPosition(node0.qp, posQuantBits, point);
point = invQuantPosition(node0.qp, posQuantBitMasks, point);
pointCloud[processedPointCount++] =
Vec3<double>(point[0], point[1], point[2]);
}
......@@ -784,7 +794,7 @@ decodeGeometryOctree(
int quantRemovedBits = (node.qp - 4) / 6;
for (int k = 0; k < 3; k++)
node.pos[k] <<= nodeSizeLog2[k] - quantRemovedBits;
node.pos = invQuantPosition(node.qp, posQuantBits, node.pos);
node.pos = invQuantPosition(node.qp, posQuantBitMasks, node.pos);
}
*nodesRemaining = std::move(fifo);
return;
......
......@@ -487,17 +487,19 @@ calculateNodeQps(int baseQp, It nodesBegin, It nodesEnd)
void
geometryQuantization(
PCCPointSet3& pointCloud, PCCOctree3Node& node, int nodeSizeLog2)
PCCPointSet3& pointCloud, PCCOctree3Node& node, Vec3<int> nodeSizeLog2)
{
QuantizerGeom quantizer = QuantizerGeom(node.qp);
int qpShift = (node.qp - 4) / 6;
int quantBitsMask = (1 << nodeSizeLog2) - 1;
Vec3<uint32_t> end_pos_quant = ((1 << nodeSizeLog2) >> qpShift) - 1;
for (int i = node.start; i < node.end; i++) {
for (int k = 0; k < 3; k++) {
for (int k = 0; k < 3; k++) {
int quantBitsMask = (1 << nodeSizeLog2[k]) - 1;
uint32_t clipMax = ((1 << nodeSizeLog2[k]) >> qpShift) - 1;
for (int i = node.start; i < node.end; i++) {
uint32_t pos = uint32_t(pointCloud[i][k]);
uint32_t quantPos = quantizer.quantize(pos & quantBitsMask);
quantPos = PCCClip(quantPos, 0, end_pos_quant[k]);
quantPos = PCCClip(quantPos, 0, clipMax);
// NB: this representation is: |pppppp00qqq|, which isn't the
// same used by the decoder: (|ppppppqqq|00)
......@@ -510,13 +512,14 @@ geometryQuantization(
void
geometryScale(
PCCPointSet3& pointCloud, PCCOctree3Node& node, int quantNodeMaxDimLog2)
PCCPointSet3& pointCloud, PCCOctree3Node& node, Vec3<int> quantNodeSizeLog2)
{
QuantizerGeom quantizer = QuantizerGeom(node.qp);
int qpShift = (node.qp - 4) / 6;
int quantBitsMask = (1 << quantNodeMaxDimLog2) - 1;
for (int i = node.start; i < node.end; i++) {
for (int k = 0; k < 3; k++) {
for (int k = 0; k < 3; k++) {
int quantBitsMask = (1 << quantNodeSizeLog2[k]) - 1;
for (int i = node.start; i < node.end; i++) {
uint32_t pos = uint32_t(pointCloud[i][k]);
uint32_t quantPos = pos & quantBitsMask;
pointCloud[i][k] = (pos & ~quantBitsMask) | quantizer.scale(quantPos);
......@@ -649,7 +652,7 @@ encodeGeometryOctree(
Vec3<uint32_t> occupancyAtlasOrigin(0xffffffff);
// the node size where quantisation is performed
int quantNodeMaxDimLog2 = 0;
Vec3<int> quantNodeSizeLog2 = 0;
int numLvlsUntilQuantization = -1;
if (gps.geom_scaling_enabled_flag) {
numLvlsUntilQuantization = 0;
......@@ -661,7 +664,7 @@ encodeGeometryOctree(
// applied to the root node, just set the node qp
if (numLvlsUntilQuantization == 0) {
quantNodeMaxDimLog2 = nodeMaxDimLog2;
quantNodeSizeLog2 = nodeSizeLog2;
node00.qp = sliceQp;
}
......@@ -694,7 +697,6 @@ encodeGeometryOctree(
nodeSizeLog2, maxNumImplicitQtbtBeforeOt, minSizeImplicitQtbt);
pointSortMask = qtBtChildSize(nodeSizeLog2, childSizeLog2);
occupancySkip = nonSplitQtBtAxes(nodeSizeLog2, childSizeLog2);
nodeMaxDimLog2--;
encoder.beginOctreeLevel();
......@@ -706,7 +708,7 @@ encodeGeometryOctree(
// determing a per node QP at the appropriate level
// NB: this has no effect here if geom_octree_qp_offset_depth=0
if (--numLvlsUntilQuantization == 0) {
quantNodeMaxDimLog2 = nodeMaxDimLog2;
quantNodeSizeLog2 = nodeSizeLog2;
calculateNodeQps(gps.geom_base_qp, fifo.begin(), fifoCurrLvlEnd);
}
}
......@@ -723,8 +725,18 @@ encodeGeometryOctree(
auto effectiveNodeSizeLog2 = nodeSizeLog2 - shiftBits;
auto effectiveChildSizeLog2 = childSizeLog2 - shiftBits;
// todo(??): the following needs to be reviewed, it is added to make
// quantisation work with qtbt.
Vec3<int> actualNodeSizeLog2, actualChildSizeLog2;
for (int k = 0; k < 3; k++) {
actualNodeSizeLog2[k] = std::max(nodeSizeLog2[k], shiftBits);
actualChildSizeLog2[k] = std::max(childSizeLog2[k], shiftBits);
}
// todo(??): atlasShift may be wrong too
occupancySkip = nonSplitQtBtAxes(actualNodeSizeLog2, actualChildSizeLog2);
if (numLvlsUntilQuantization == 0) {
geometryQuantization(pointCloud, node0, quantNodeMaxDimLog2);
geometryQuantization(pointCloud, node0, quantNodeSizeLog2);
if (gps.geom_unique_points_flag)
checkDuplicatePoints(pointCloud, node0, pointIdxToDmIdx);
}
......@@ -805,7 +817,7 @@ encodeGeometryOctree(
int childStart = node0.start;
// inverse quantise any quantised positions
geometryScale(pointCloud, node0, quantNodeMaxDimLog2);
geometryScale(pointCloud, node0, quantNodeSizeLog2);
for (int i = 0; i < 8; i++) {
if (!childCounts[i]) {
......@@ -875,7 +887,7 @@ encodeGeometryOctree(
if (directModeUsed) {
// inverse quantise any quantised positions
geometryScale(pointCloud, node0, quantNodeMaxDimLog2);
geometryScale(pointCloud, node0, quantNodeSizeLog2);
// point reordering to match decoder's order
for (auto idx = child.start; idx < child.end; idx++)
......@@ -912,7 +924,7 @@ encodeGeometryOctree(
for (auto& node : fifo) {
for (int k = 0; k < 3; k++)
node.pos[k] <<= nodeSizeLog2[k];
geometryScale(pointCloud, node, quantNodeMaxDimLog2);
geometryScale(pointCloud, node, quantNodeSizeLog2);
}
*nodesRemaining = std::move(fifo);
return;
......
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment