Method setDimensionality

filters/CovarianceFeaturesFilter.cpp:204–367

}

void CovarianceFeaturesFilter::setDimensionality(PointView &view, const PointId &id, const KD3Index &kdi)
{
    using namespace Eigen;
    
    PointRef p = view.point(id);

    // find neighbors, either by radius or k nearest neighbors
    PointIdList ids;
    if (m_optimal)
    {
        ids = kdi.neighbors(p, p.getFieldAs<uint64_t>(Id::OptimalKNN), 1);
    }
    else if (m_radiusArg->set())
    {
        ids = kdi.radius(p, m_radius);
        if (ids.size() < (size_t)m_minK) {
            log()->get(LogLevel::Info)
                << "Skipping point " << id << ". Found " << ids.size()
                << " neighbors but required " << m_minK << ".\n";
            return;
        }
    }
    else
    {
        ids = kdi.neighbors(p, m_knn + 1, m_stride);
    }

    // compute covariance of the neighborhood
    auto B = math::computeCovariance(view, ids);

    // Check if the covariance matrix is all zeros
    if (B.isZero())
    {
        log()->get(LogLevel::Info)
            << "Skipping point " << id
            << ". Covariance matrix is all zeros. This suggests a large number "
               "of redundant points. Consider using filters.sample with a "
               "small radius to remove redundant points.\n";
        return;
    }

    // perform the eigen decomposition
    SelfAdjointEigenSolver<Matrix3d> solver(B);
    if (solver.info() != Success)
        throwError("Cannot perform eigen decomposition.");

    // Extract eigenvalues and eigenvectors in decreasing order (largest eigenvalue first)
    auto ev = solver.eigenvalues();
    std::vector<double> lambda = {((std::max)(ev[2],0.0)),
                                  ((std::max)(ev[1],0.0)),
                                  ((std::max)(ev[0],0.0))};
    double sum = std::accumulate(lambda.begin(), lambda.end(), 0.0);

    if (lambda[0] == 0)
        throwError("Eigenvalues are all 0. Can't compute local features.");

    if (m_mode == Mode::SQRT)
    {