void EstimateRankFilter::filter(PointView& view)
{
KD3Index& kdi = view.build3dIndex();
for (PointId i = 0; i < view.size(); ++i)
{
// find the k-nearest neighbors
auto ids = kdi.neighbors(i, m_knn);
view.setField(m_rank, i, eigen::computeRank(view, ids, m_thresh));
}
}
void NormalFilter::filter(PointView& view)
{
KD3Index& kdi = view.build3dIndex();
for (PointId i = 0; i < view.size(); ++i)
{
// find the k-nearest neighbors
auto ids = kdi.neighbors(i, m_knn);
// compute covariance of the neighborhood
auto B = eigen::computeCovariance(view, ids);
// perform the eigen decomposition
Eigen::SelfAdjointEigenSolver<Eigen::Matrix3f> solver(B);
if (solver.info() != Eigen::Success)
throwError("Cannot perform eigen decomposition.");
auto eval = solver.eigenvalues();
Eigen::Vector3f normal = solver.eigenvectors().col(0);
if (m_viewpointArg->set())
{
PointRef p = view.point(i);
Eigen::Vector3f vp(
m_viewpoint.x - p.getFieldAs<double>(Dimension::Id::X),
m_viewpoint.y - p.getFieldAs<double>(Dimension::Id::Y),
m_viewpoint.z - p.getFieldAs<double>(Dimension::Id::Z));
if (vp.dot(normal) < 0)
normal *= -1.0;
}
else if (m_up)
{
if (normal[2] < 0)
normal *= -1.0;
}
view.setField(Dimension::Id::NormalX, i, normal[0]);
view.setField(Dimension::Id::NormalY, i, normal[1]);
view.setField(Dimension::Id::NormalZ, i, normal[2]);
double sum = eval[0] + eval[1] + eval[2];
view.setField(Dimension::Id::Curvature, i,
sum ? std::fabs(eval[0] / sum) : 0);
}
}
void LOFFilter::filter(PointView& view)
{
using namespace Dimension;
KD3Index& index = view.build3dIndex();
// Increment the minimum number of points, as knnSearch will be returning
// the neighbors along with the query point.
m_minpts++;
// First pass: Compute the k-distance for each point.
// The k-distance is the Euclidean distance to k-th nearest neighbor.
log()->get(LogLevel::Debug) << "Computing k-distances...\n";
for (PointId i = 0; i < view.size(); ++i)
{
std::vector<PointId> indices(m_minpts);
std::vector<double> sqr_dists(m_minpts);
index.knnSearch(i, m_minpts, &indices, &sqr_dists);
view.setField(m_kdist, i, std::sqrt(sqr_dists[m_minpts-1]));
}
// Second pass: Compute the local reachability distance for each point.
// For each neighbor point, the reachability distance is the maximum value
// of that neighbor's k-distance and the distance between the neighbor and
// the current point. The lrd is the inverse of the mean of the reachability
// distances.
log()->get(LogLevel::Debug) << "Computing lrd...\n";
for (PointId i = 0; i < view.size(); ++i)
{
std::vector<PointId> indices(m_minpts);
std::vector<double> sqr_dists(m_minpts);
index.knnSearch(i, m_minpts, &indices, &sqr_dists);
double M1 = 0.0;
point_count_t n = 0;
for (PointId j = 0; j < indices.size(); ++j)
{
double k = view.getFieldAs<double>(m_kdist, indices[j]);
double reachdist = (std::max)(k, std::sqrt(sqr_dists[j]));
M1 += (reachdist - M1) / ++n;
}
view.setField(m_lrd, i, 1.0 / M1);
}
// Third pass: Compute the local outlier factor for each point.
// The LOF is the average of the lrd's for a neighborhood of points.
log()->get(LogLevel::Debug) << "Computing LOF...\n";
for (PointId i = 0; i < view.size(); ++i)
{
double lrdp = view.getFieldAs<double>(m_lrd, i);
std::vector<PointId> indices(m_minpts);
std::vector<double> sqr_dists(m_minpts);
index.knnSearch(i, m_minpts, &indices, &sqr_dists);
double M1 = 0.0;
point_count_t n = 0;
for (auto const& j : indices)
{
M1 += (view.getFieldAs<double>(m_lrd, j) / lrdp - M1) / ++n;
}
view.setField(m_lof, i, M1);
}
}
