void DataPointsFiltersImpl<T>::MaxQuantileOnAxisDataPointsFilter::inPlaceFilter(
DataPoints& cloud)
{
if (int(dim) >= cloud.features.rows())
throw InvalidParameter((boost::format("MaxQuantileOnAxisDataPointsFilter: Error, filtering on dimension number %1%, larger than feature dimensionality %2%") % dim % cloud.features.rows()).str());
const int nbPointsIn = cloud.features.cols();
const int nbPointsOut = nbPointsIn * ratio;
// build array
vector<T> values;
values.reserve(cloud.features.cols());
for (int x = 0; x < cloud.features.cols(); ++x)
values.push_back(cloud.features(dim, x));
// get quartiles value
nth_element(values.begin(), values.begin() + (values.size() * ratio), values.end());
const T limit = values[nbPointsOut];
// copy towards beginning the elements we keep
int j = 0;
for (int i = 0; i < nbPointsIn; i++)
{
if (cloud.features(dim, i) < limit)
{
assert(j <= i);
cloud.setColFrom(j, cloud, i);
j++;
}
}
assert(j <= nbPointsOut);
cloud.conservativeResize(j);
}
void DataPointsFiltersImpl<T>::MaxDistDataPointsFilter::inPlaceFilter(
DataPoints& cloud)
{
if (dim >= cloud.features.rows() - 1)
{
throw InvalidParameter(
(boost::format("MaxDistDataPointsFilter: Error, filtering on dimension number %1%, larger than authorized axis id %2%") % dim % (cloud.features.rows() - 2)).str());
}
const int nbPointsIn = cloud.features.cols();
const int nbRows = cloud.features.rows();
int j = 0;
if(dim == -1) // Euclidean distance
{
for (int i = 0; i < nbPointsIn; i++)
{
const T absMaxDist = anyabs(maxDist);
if (cloud.features.col(i).head(nbRows-1).norm() < absMaxDist)
{
cloud.setColFrom(j, cloud, i);
j++;
}
}
}
else // Single-axis distance
{
for (int i = 0; i < nbPointsIn; i++)
{
if ((cloud.features(dim, i)) < maxDist)
{
cloud.setColFrom(j, cloud, i);
j++;
}
}
}
cloud.conservativeResize(j);
}
void CovarianceSamplingDataPointsFilter<T>::inPlaceFilter(DataPoints& cloud)
{
const std::size_t featDim(cloud.features.rows());
assert(featDim == 4); //3D pts only
//Check number of points
const std::size_t nbPoints = cloud.getNbPoints();
if(nbSample >= nbPoints)
return;
//Check if there is normals info
if (!cloud.descriptorExists("normals"))
throw InvalidField("OrientNormalsDataPointsFilter: Error, cannot find normals in descriptors.");
const auto& normals = cloud.getDescriptorViewByName("normals");
std::vector<std::size_t> keepIndexes;
keepIndexes.resize(nbSample);
///---- Part A, as we compare the cloud with himself, the overlap is 100%, so we keep all points
//A.1 and A.2 - Compute candidates
std::vector<std::size_t> candidates ;
candidates.resize(nbPoints);
for (std::size_t i = 0; i < nbPoints; ++i) candidates[i] = i;
const std::size_t nbCandidates = candidates.size();
//Compute centroid
Vector3 center;
for(std::size_t i = 0; i < featDim - 1; ++i) center(i) = T(0.);
for (std::size_t i = 0; i < nbCandidates; ++i)
for (std::size_t f = 0; f <= 3; ++f)
center(f) += cloud.features(f,candidates[i]);
for(std::size_t i = 0; i <= 3; ++i) center(i) /= T(nbCandidates);
//Compute torque normalization
T Lnorm = 1.0;
if(normalizationMethod == TorqueNormMethod::L1)
{
Lnorm = 1.0;
}
else if(normalizationMethod == TorqueNormMethod::Lavg)
{
Lnorm = 0.0;
for (std::size_t i = 0; i < nbCandidates; ++i)
Lnorm += (cloud.features.col(candidates[i]).head(3) - center).norm();
Lnorm /= nbCandidates;
}
else if(normalizationMethod == TorqueNormMethod::Lmax)
{
const Vector minValues = cloud.features.rowwise().minCoeff();
const Vector maxValues = cloud.features.rowwise().maxCoeff();
const Vector3 radii = maxValues.head(3) - minValues.head(3);
Lnorm = radii.maxCoeff() / 2.; //radii.mean() / 2.;
}
//A.3 - Compute 6x6 covariance matrix + EigenVectors
auto computeCovariance = [Lnorm, nbCandidates, &cloud, ¢er, &normals, &candidates](Matrix66 & cov) -> void {
//Compute F matrix, see Eq. (4)
Eigen::Matrix<T, 6, Eigen::Dynamic> F(6, nbCandidates);
for(std::size_t i = 0; i < nbCandidates; ++i)
{
const Vector3 p = cloud.features.col(candidates[i]).head(3) - center; // pi-c
const Vector3 ni = normals.col(candidates[i]).head(3);
//compute (1 / L) * (pi - c) x ni
F.template block<3, 1>(0, i) = (1. / Lnorm) * p.cross(ni);
//set ni part
F.template block<3, 1>(3, i) = ni;
}
// Compute the covariance matrix Cov = FF'
cov = F * F.transpose();
};
Matrix66 covariance;
computeCovariance(covariance);
Eigen::EigenSolver<Matrix66> solver(covariance);
const Matrix66 eigenVe = solver.eigenvectors().real();
const Vector6 eigenVa = solver.eigenvalues().real();
///---- Part B
//B.1 - Compute the v-6 for each candidate
std::vector<Vector6, Eigen::aligned_allocator<Vector6>> v; // v[i] = [(pi-c) x ni ; ni ]'
v.resize(nbCandidates);
for(std::size_t i = 0; i < nbCandidates; ++i)
{
const Vector3 p = cloud.features.col(candidates[i]).head(3) - center; // pi-c
const Vector3 ni = normals.col(candidates[i]).head(3);
v[i].template block<3, 1>(0, 0) = (1. / Lnorm) * p.cross(ni);
v[i].template block<3, 1>(3, 0) = ni;
}
//B.2 - Compute the 6 sorted list based on dot product (vi . Xk) = magnitude, with Xk the kth-EigenVector
std::vector<std::list<std::pair<int, T>>> L; // contain list of pair (index, magnitude) contribution to the eigens vectors
L.resize(6);
//sort by decreasing magnitude
auto comp = [](const std::pair<int, T>& p1, const std::pair<int, T>& p2) -> bool {
return p1.second > p2.second;
};
for(std::size_t k = 0; k < 6; ++k)
{
for(std::size_t i = 0; i < nbCandidates; ++i )
{
L[k].push_back(std::make_pair(i, std::fabs( v[i].dot(eigenVe.template block<6,1>(0, k)) )));
}
L[k].sort(comp);
}
std::vector<T> t(6, T(0.)); //contains the sums of squared magnitudes
std::vector<bool> sampledPoints(nbCandidates, false); //maintain flag to avoid resampling the same point in an other list
///Add point iteratively till we got the desired number of point
for(std::size_t i = 0; i < nbSample; ++i)
{
//B.3 - Equally constrained all eigen vectors
// magnitude contribute to t_i where i is the indice of th least contrained eigen vector
//Find least constrained vector
std::size_t k = 0;
for (std::size_t i = 0; i < 6; ++i)
{
if (t[k] > t[i])
k = i;
}
// Add the point from the top of the list corresponding to the dimension to the set of samples
while(sampledPoints[L[k].front().first])
L[k].pop_front(); //remove already sampled point
//Get index to keep
const std::size_t idToKeep = static_cast<std::size_t>(L[k].front().first);
L[k].pop_front();
sampledPoints[idToKeep] = true; //set flag to avoid resampling
//B.4 - Update the running total
for (std::size_t k = 0; k < 6; ++k)
{
const T magnitude = v[idToKeep].dot(eigenVe.template block<6, 1>(0, k));
t[k] += (magnitude * magnitude);
}
keepIndexes[i] = candidates[idToKeep];
}
//TODO: evaluate performances between this solution and sorting the indexes
// We build map of (old index to new index), in case we sample pts at the begining of the pointcloud
std::unordered_map<std::size_t, std::size_t> mapidx;
std::size_t idx = 0;
///(4) Sample the point cloud
for(std::size_t id : keepIndexes)
{
//retrieve index from lookup table if sampling in already switched element
if(id<idx)
id = mapidx[id];
//Switch columns id and idx
cloud.swapCols(idx, id);
//Maintain new index position
mapidx[idx] = id;
//Update index
++idx;
}
cloud.conservativeResize(nbSample);
}
void DataPointsFiltersImpl<T>::VoxelGridDataPointsFilter::inPlaceFilter(DataPoints& cloud)
{
const int numPoints(cloud.features.cols());
const int featDim(cloud.features.rows());
const int descDim(cloud.descriptors.rows());
assert (featDim == 3 || featDim == 4);
int insertDim(0);
if (averageExistingDescriptors)
{
// TODO: this should be in the form of an assert
// Validate descriptors and labels
for(unsigned int i = 0; i < cloud.descriptorLabels.size(); i++)
insertDim += cloud.descriptorLabels[i].span;
if (insertDim != descDim)
throw InvalidField("VoxelGridDataPointsFilter: Error, descriptor labels do not match descriptor data");
}
// TODO: Check that the voxel size is not too small, given the size of the data
// Calculate number of divisions along each axis
Vector minValues = cloud.features.rowwise().minCoeff();
Vector maxValues = cloud.features.rowwise().maxCoeff();
T minBoundX = minValues.x() / vSizeX;
T maxBoundX = maxValues.x() / vSizeX;
T minBoundY = minValues.y() / vSizeY;
T maxBoundY = maxValues.y() / vSizeY;
T minBoundZ = 0;
T maxBoundZ = 0;
if (featDim == 4)
{
minBoundZ = minValues.z() / vSizeZ;
maxBoundZ = maxValues.z() / vSizeZ;
}
// number of divisions is total size / voxel size voxels of equal length + 1
// with remaining space
unsigned int numDivX = 1 + maxBoundX - minBoundX;
unsigned int numDivY = 1 + maxBoundY - minBoundY;;
unsigned int numDivZ = 0;
// If a 3D point cloud
if (featDim == 4 )
numDivZ = 1 + maxBoundZ - minBoundZ;
unsigned int numVox = numDivX * numDivY;
if ( featDim == 4)
numVox *= numDivZ;
// Assume point cloud is randomly ordered
// compute a linear index of the following type
// i, j, k are the component indices
// nx, ny number of divisions in x and y components
// idx = i + j * nx + k * nx * ny
std::vector<unsigned int> indices(numPoints);
// vector to hold the first point in a voxel
// this point will be ovewritten in the input cloud with
// the output value
std::vector<Voxel>* voxels;
// try allocating vector. If too big return error
try {
voxels = new std::vector<Voxel>(numVox);
} catch (std::bad_alloc&) {
throw InvalidParameter((boost::format("VoxelGridDataPointsFilter: Memory allocation error with %1% voxels. Try increasing the voxel dimensions.") % numVox).str());
}
for (int p = 0; p < numPoints; p++ )
{
unsigned int i = floor(cloud.features(0,p)/vSizeX - minBoundX);
unsigned int j = floor(cloud.features(1,p)/vSizeY- minBoundY);
unsigned int k = 0;
unsigned int idx;
if ( featDim == 4 )
{
k = floor(cloud.features(2,p)/vSizeZ - minBoundZ);
idx = i + j * numDivX + k * numDivX * numDivY;
}
else
{
idx = i + j * numDivX;
}
unsigned int pointsInVox = (*voxels)[idx].numPoints + 1;
if (pointsInVox == 1)
{
(*voxels)[idx].firstPoint = p;
}
(*voxels)[idx].numPoints = pointsInVox;
indices[p] = idx;
}
// store which points contain voxel position
std::vector<unsigned int> pointsToKeep;
// Store voxel centroid in output
if (useCentroid)
{
// Iterate through the indices and sum values to compute centroid
for (int p = 0; p < numPoints ; p++)
{
unsigned int idx = indices[p];
unsigned int firstPoint = (*voxels)[idx].firstPoint;
// If this is the first point in the voxel, leave as is
// if not sum up this point for centroid calculation
if (firstPoint != p)
{
// Sum up features and descriptors (if we are also averaging descriptors)
for (int f = 0; f < (featDim - 1); f++ )
{
cloud.features(f,firstPoint) += cloud.features(f,p);
}
if (averageExistingDescriptors) {
for (int d = 0; d < descDim; d++)
{
cloud.descriptors(d,firstPoint) += cloud.descriptors(d,p);
}
}
}
}
// Now iterating through the voxels
// Normalize sums to get centroid (average)
// Some voxels may be empty and are discarded
for( int idx = 0; idx < numVox; idx++)
{
unsigned int numPoints = (*voxels)[idx].numPoints;
unsigned int firstPoint = (*voxels)[idx].firstPoint;
if(numPoints > 0)
{
for ( int f = 0; f < (featDim - 1); f++ )
cloud.features(f,firstPoint) /= numPoints;
if (averageExistingDescriptors) {
for ( int d = 0; d < descDim; d++ )
cloud.descriptors(d,firstPoint) /= numPoints;
}
pointsToKeep.push_back(firstPoint);
}
}
}
else
{
// Although we don't sum over the features, we may still need to sum the descriptors
if (averageExistingDescriptors)
{
// Iterate through the indices and sum values to compute centroid
for (int p = 0; p < numPoints ; p++)
{
unsigned int idx = indices[p];
unsigned int firstPoint = (*voxels)[idx].firstPoint;
// If this is the first point in the voxel, leave as is
// if not sum up this point for centroid calculation
if (firstPoint != p)
{
for (int d = 0; d < descDim; d++ )
{
cloud.descriptors(d,firstPoint) += cloud.descriptors(d,p);
}
}
}
}
for (int idx = 0; idx < numVox; idx++)
{
unsigned int numPoints = (*voxels)[idx].numPoints;
unsigned int firstPoint = (*voxels)[idx].firstPoint;
if (numPoints > 0)
{
// get back voxel indices in grid format
// If we are in the last division, the voxel is smaller in size
// We adjust the center as from the end of the last voxel to the bounding area
unsigned int i = 0;
unsigned int j = 0;
unsigned int k = 0;
if (featDim == 4)
{
k = idx / (numDivX * numDivY);
if (k == numDivZ)
cloud.features(3,firstPoint) = maxValues.z() - (k-1) * vSizeZ/2;
else
cloud.features(3,firstPoint) = k * vSizeZ + vSizeZ/2;
}
j = (idx - k * numDivX * numDivY) / numDivX;
if (j == numDivY)
cloud.features(2,firstPoint) = maxValues.y() - (j-1) * vSizeY/2;
else
cloud.features(2,firstPoint) = j * vSizeY + vSizeY / 2;
i = idx - k * numDivX * numDivY - j * numDivX;
if (i == numDivX)
cloud.features(1,firstPoint) = maxValues.x() - (i-1) * vSizeX/2;
else
cloud.features(1,firstPoint) = i * vSizeX + vSizeX / 2;
// Descriptors : normalize if we are averaging or keep as is
if (averageExistingDescriptors) {
for ( int d = 0; d < descDim; d++ )
cloud.descriptors(d,firstPoint) /= numPoints;
}
pointsToKeep.push_back(firstPoint);
}
}
}
// deallocate memory for voxels information
delete voxels;
// Move the points to be kept to the start
// Bring the data we keep to the front of the arrays then
// wipe the leftover unused space.
std::sort(pointsToKeep.begin(), pointsToKeep.end());
int numPtsOut = pointsToKeep.size();
for (int i = 0; i < numPtsOut; i++){
int k = pointsToKeep[i];
assert(i <= k);
cloud.features.col(i) = cloud.features.col(k);
if (cloud.descriptors.rows() != 0)
cloud.descriptors.col(i) = cloud.descriptors.col(k);
}
cloud.features.conservativeResize(Eigen::NoChange, numPtsOut);
if (cloud.descriptors.rows() != 0)
cloud.descriptors.conservativeResize(Eigen::NoChange, numPtsOut);
}
