KDTreeCapable.h

Go to the documentation of this file.

/* +------------------------------------------------------------------------+
   |                     Mobile Robot Programming Toolkit (MRPT)            |
   |                          https://www.mrpt.org/                         |
   |                                                                        |
   | Copyright (c) 2005-2019, Individual contributors, see AUTHORS file     |
   | See: https://www.mrpt.org/Authors - All rights reserved.               |
   | Released under BSD License. See: https://www.mrpt.org/License          |
   +------------------------------------------------------------------------+ */
#pragma once

// nanoflann library:
#include <mrpt/core/common.h>
#include <mrpt/core/exceptions.h>
#include <mrpt/math/TPoint2D.h>
#include <mrpt/math/TPoint3D.h>
#include <array>
#include <atomic>
#include <memory>  // unique_ptr
#include <mutex>
#include <nanoflann.hpp>

namespace mrpt::math
{
/** \addtogroup kdtree_grp KD-Trees
 *  \ingroup mrpt_math_grp
 *  @{ */

/** A generic adaptor class for providing Nearest Neighbor (NN) lookup via the
 * `nanoflann` library.
 *   This makes use of the CRTP design pattern.
 *
 *  Derived classes must be aware of the need to call
 * "kdtree_mark_as_outdated()" when the data points
 *   change to mark the cached KD-tree (an "index") as invalid, and also
 * implement the following interface
 *   (note that these are not virtual functions due to the usage of CRTP):
 *
 *  \code
 *   // Must return the number of data points
 *   inline size_t kdtree_get_point_count() const { ... }
 *
 *   // Returns the distance between the vector "p1[0:size-1]" and the data
 * point with index "idx_p2" stored in the class:
 *   inline float kdtree_distance(const float *p1, const size_t idx_p2,size_t
 * size) const { ... }
 *
 *   // Returns the dim'th component of the idx'th point in the class:
 *   inline num_t kdtree_get_pt(const size_t idx, int dim) const { ... }
 *
 *   // Optional bounding-box computation: return false to default to a standard
 * bbox computation loop.
 *   //   Return true if the BBOX was already computed by the class and returned
 * in "bb" so it can be avoided to redo it again.
 *   //   Look at bb.size() to find out the expected dimensionality (e.g. 2 or 3
 * for point clouds)
 *   template <class BBOX>
 *   bool kdtree_get_bbox(BBOX &bb) const
 *   {
 *      bb[0].low = ...; bb[0].high = ...;  // "x" limits
 *      return true;
 *   }
 *
 *  \endcode
 *
 * The KD-tree index will be built on demand only upon call of any of the query
 * methods provided by this class.
 *
 *  Notice that there is only ONE internal cached KD-tree, so if a method to
 * query a 2D point is called,
 *  then another method for 3D points, then again the 2D method, three KD-trees
 * will be built. So, try
 *  to group all the calls for a given dimensionality together or build
 * different class instances for
 *  queries of each dimensionality, etc.
 *
 *  \sa See some of the derived classes for example implementations. See also
 * the documentation of nanoflann
 * \ingroup mrpt_math_grp
 */
template <
    class Derived, typename num_t = float,
    typename metric_t = nanoflann::L2_Simple_Adaptor<num_t, Derived>>
class KDTreeCapable
{
   public:
    // Types ---------------
    using self_t = KDTreeCapable<Derived, num_t, metric_t>;
    // ---------------------

    /// Constructor
    inline KDTreeCapable() = default;
    KDTreeCapable(const KDTreeCapable&) : KDTreeCapable() {}
    KDTreeCapable& operator=(const KDTreeCapable&)
    {
        kdtree_mark_as_outdated();
        return *this;
    }

    /// CRTP helper method
    inline const Derived& derived() const
    {
        return *static_cast<const Derived*>(this);
    }
    /// CRTP helper method
    inline Derived& derived() { return *static_cast<Derived*>(this); }
    struct TKDTreeSearchParams
    {
        TKDTreeSearchParams() = default;
        /** Max points per leaf */
        size_t leaf_max_size{10};
    };

    /** Parameters to tune the ANN searches */
    TKDTreeSearchParams kdtree_search_params;

    /** @name Public utility methods to query the KD-tree
        @{ */

    /** KD Tree-based search for the closest point (only ONE) to some given 2D
     *coordinates.
     *  This method automatically build the "m_kdtree_data" structure when:
     *      - It is called for the first time
     *      - The map has changed
     *      - The KD-tree was build for 3D.
     *
     * \param x0  The X coordinate of the query.
     * \param y0  The Y coordinate of the query.
     * \param out_x The X coordinate of the found closest correspondence.
     * \param out_y The Y coordinate of the found closest correspondence.
     * \param out_dist_sqr The square distance between the query and the
     *returned point.
     *
     * \return The index of the closest point in the map array.
     *  \sa kdTreeClosestPoint3D, kdTreeTwoClosestPoint2D
     */
    inline size_t kdTreeClosestPoint2D(
        float x0, float y0, float& out_x, float& out_y,
        float& out_dist_sqr) const
    {
        MRPT_START
        rebuild_kdTree_2D();  // First: Create the 2D KD-Tree if required
        if (!m_kdtree2d_data.m_num_points)
            THROW_EXCEPTION("There are no points in the KD-tree.");

        const size_t knn = 1;  // Number of points to retrieve
        size_t ret_index;
        nanoflann::KNNResultSet<num_t> resultSet(knn);
        resultSet.init(&ret_index, &out_dist_sqr);

        const std::array<num_t, 2> query_point{{x0, y0}};
        m_kdtree2d_data.index->findNeighbors(
            resultSet, &query_point[0], nanoflann::SearchParams());

        // Copy output to user vars:
        out_x = derived().kdtree_get_pt(ret_index, 0);
        out_y = derived().kdtree_get_pt(ret_index, 1);

        return ret_index;
        MRPT_END
    }

    /// \overload
    inline size_t kdTreeClosestPoint2D(
        float x0, float y0, float& out_dist_sqr) const
    {
        MRPT_START
        rebuild_kdTree_2D();  // First: Create the 2D KD-Tree if required
        if (!m_kdtree2d_data.m_num_points)
            THROW_EXCEPTION("There are no points in the KD-tree.");

        const size_t knn = 1;  // Number of points to retrieve
        size_t ret_index;
        nanoflann::KNNResultSet<num_t> resultSet(knn);
        resultSet.init(&ret_index, &out_dist_sqr);

        const std::array<num_t, 2> query_point{{x0, y0}};
        m_kdtree2d_data.index->findNeighbors(
            resultSet, &query_point[0], nanoflann::SearchParams());

        return ret_index;
        MRPT_END
    }

    /// \overload
    inline size_t kdTreeClosestPoint2D(
        const TPoint2D& p0, TPoint2D& pOut, float& outDistSqr) const
    {
        float dmy1, dmy2;
        size_t res = kdTreeClosestPoint2D(
            static_cast<float>(p0.x), static_cast<float>(p0.y), dmy1, dmy2,
            outDistSqr);
        pOut.x = dmy1;
        pOut.y = dmy2;
        return res;
    }

    /** Like kdTreeClosestPoint2D, but just return the square error from some
     * point to its closest neighbor.
     */
    inline float kdTreeClosestPoint2DsqrError(float x0, float y0) const
    {
        float closerx, closery, closer_dist;
        kdTreeClosestPoint2D(x0, y0, closerx, closery, closer_dist);
        return closer_dist;
    }

    inline float kdTreeClosestPoint2DsqrError(const TPoint2D& p0) const
    {
        return kdTreeClosestPoint2DsqrError(
            static_cast<float>(p0.x), static_cast<float>(p0.y));
    }

    /** KD Tree-based search for the TWO closest point to some given 2D
     *coordinates.
     *  This method automatically build the "m_kdtree_data" structure when:
     *      - It is called for the first time
     *      - The map has changed
     *      - The KD-tree was build for 3D.
     *
     * \param x0  The X coordinate of the query.
     * \param y0  The Y coordinate of the query.
     * \param out_x1 The X coordinate of the first correspondence.
     * \param out_y1 The Y coordinate of the first correspondence.
     * \param out_x2 The X coordinate of the second correspondence.
     * \param out_y2 The Y coordinate of the second correspondence.
     * \param out_dist_sqr1 The square distance between the query and the first
     *returned point.
     * \param out_dist_sqr2 The square distance between the query and the
     *second returned point.
     *
     *  \sa kdTreeClosestPoint2D
     */
    inline void kdTreeTwoClosestPoint2D(
        float x0, float y0, float& out_x1, float& out_y1, float& out_x2,
        float& out_y2, float& out_dist_sqr1, float& out_dist_sqr2) const
    {
        MRPT_START
        rebuild_kdTree_2D();  // First: Create the 2D KD-Tree if required
        if (!m_kdtree2d_data.m_num_points)
            THROW_EXCEPTION("There are no points in the KD-tree.");

        const size_t knn = 2;  // Number of points to retrieve
        size_t ret_indexes[2];
        float ret_sqdist[2];
        nanoflann::KNNResultSet<num_t> resultSet(knn);
        resultSet.init(&ret_indexes[0], &ret_sqdist[0]);

        const std::array<num_t, 2> query_point{{x0, y0}};
        m_kdtree2d_data.index->findNeighbors(
            resultSet, &query_point[0], nanoflann::SearchParams());

        // Copy output to user vars:
        out_x1 = derived().kdtree_get_pt(ret_indexes[0], 0);
        out_y1 = derived().kdtree_get_pt(ret_indexes[0], 1);
        out_dist_sqr1 = ret_sqdist[0];

        out_x2 = derived().kdtree_get_pt(ret_indexes[1], 0);
        out_y2 = derived().kdtree_get_pt(ret_indexes[1], 1);
        out_dist_sqr2 = ret_sqdist[0];

        MRPT_END
    }

    inline void kdTreeTwoClosestPoint2D(
        const TPoint2D& p0, TPoint2D& pOut1, TPoint2D& pOut2,
        float& outDistSqr1, float& outDistSqr2) const
    {
        float dmy1, dmy2, dmy3, dmy4;
        kdTreeTwoClosestPoint2D(
            p0.x, p0.y, dmy1, dmy2, dmy3, dmy4, outDistSqr1, outDistSqr2);
        pOut1.x = static_cast<double>(dmy1);
        pOut1.y = static_cast<double>(dmy2);
        pOut2.x = static_cast<double>(dmy3);
        pOut2.y = static_cast<double>(dmy4);
    }

    /** KD Tree-based search for the N closest point to some given 2D
     *coordinates.
     *  This method automatically build the "m_kdtree_data" structure when:
     *      - It is called for the first time
     *      - The map has changed
     *      - The KD-tree was build for 3D.
     *
     * \param x0  The X coordinate of the query.
     * \param y0  The Y coordinate of the query.
     * \param N The number of closest points to search.
     * \param out_x The vector containing the X coordinates of the
     *correspondences.
     * \param out_y The vector containing the Y coordinates of the
     *correspondences.
     * \param out_dist_sqr The vector containing the square distance between
     *the query and the returned points.
     *
     * \return The list of indices
     *  \sa kdTreeClosestPoint2D
     *  \sa kdTreeTwoClosestPoint2D
     */
    inline std::vector<size_t> kdTreeNClosestPoint2D(
        float x0, float y0, size_t knn, std::vector<float>& out_x,
        std::vector<float>& out_y, std::vector<float>& out_dist_sqr) const
    {
        MRPT_START
        rebuild_kdTree_2D();  // First: Create the 2D KD-Tree if required
        if (!m_kdtree2d_data.m_num_points)
            THROW_EXCEPTION("There are no points in the KD-tree.");

        std::vector<size_t> ret_indexes(knn);
        out_x.resize(knn);
        out_y.resize(knn);
        out_dist_sqr.resize(knn);

        nanoflann::KNNResultSet<num_t> resultSet(knn);
        resultSet.init(&ret_indexes[0], &out_dist_sqr[0]);

        const std::array<num_t, 2> query_point{{x0, y0}};
        m_kdtree2d_data.index->findNeighbors(
            resultSet, &query_point[0], nanoflann::SearchParams());

        for (size_t i = 0; i < knn; i++)
        {
            out_x[i] = derived().kdtree_get_pt(ret_indexes[i], 0);
            out_y[i] = derived().kdtree_get_pt(ret_indexes[i], 1);
        }
        return ret_indexes;
        MRPT_END
    }

    inline std::vector<size_t> kdTreeNClosestPoint2D(
        const TPoint2D& p0, size_t N, std::vector<TPoint2D>& pOut,
        std::vector<float>& outDistSqr) const
    {
        std::vector<float> dmy1, dmy2;
        std::vector<size_t> res = kdTreeNClosestPoint2D(
            static_cast<float>(p0.x), static_cast<float>(p0.y), N, dmy1, dmy2,
            outDistSqr);
        pOut.resize(dmy1.size());
        for (size_t i = 0; i < dmy1.size(); i++)
        {
            pOut[i].x = static_cast<double>(dmy1[i]);
            pOut[i].y = static_cast<double>(dmy2[i]);
        }
        return res;
    }

    /** KD Tree-based search for the N closest point to some given 2D
     *coordinates and returns their indexes.
     *  This method automatically build the "m_kdtree_data" structure when:
     *      - It is called for the first time
     *      - The map has changed
     *      - The KD-tree was build for 3D.
     *
     * \param x0  The X coordinate of the query.
     * \param y0  The Y coordinate of the query.
     * \param N The number of closest points to search.
     * \param out_idx The indexes of the found closest correspondence.
     * \param out_dist_sqr The square distance between the query and the
     *returned point.
     *
     *  \sa kdTreeClosestPoint2D
     */
    inline void kdTreeNClosestPoint2DIdx(
        float x0, float y0, size_t knn, std::vector<size_t>& out_idx,
        std::vector<float>& out_dist_sqr) const
    {
        MRPT_START
        rebuild_kdTree_2D();  // First: Create the 2D KD-Tree if required
        if (!m_kdtree2d_data.m_num_points)
            THROW_EXCEPTION("There are no points in the KD-tree.");

        out_idx.resize(knn);
        out_dist_sqr.resize(knn);
        nanoflann::KNNResultSet<num_t> resultSet(knn);
        resultSet.init(&out_idx[0], &out_dist_sqr[0]);

        const std::array<num_t, 2> query_point{{x0, y0}};
        m_kdtree2d_data.index->findNeighbors(
            resultSet, &query_point[0], nanoflann::SearchParams());
        MRPT_END
    }

    inline void kdTreeNClosestPoint2DIdx(
        const TPoint2D& p0, size_t N, std::vector<size_t>& outIdx,
        std::vector<float>& outDistSqr) const
    {
        return kdTreeNClosestPoint2DIdx(
            static_cast<float>(p0.x), static_cast<float>(p0.y), N, outIdx,
            outDistSqr);
    }

    /** KD Tree-based search for the closest point (only ONE) to some given 3D
     *coordinates.
     *  This method automatically build the "m_kdtree_data" structure when:
     *      - It is called for the first time
     *      - The map has changed
     *      - The KD-tree was build for 2D.
     *
     * \param x0  The X coordinate of the query.
     * \param y0  The Y coordinate of the query.
     * \param z0  The Z coordinate of the query.
     * \param out_x The X coordinate of the found closest correspondence.
     * \param out_y The Y coordinate of the found closest correspondence.
     * \param out_z The Z coordinate of the found closest correspondence.
     * \param out_dist_sqr The square distance between the query and the
     *returned point.
     *
     * \return The index of the closest point in the map array.
     *  \sa kdTreeClosestPoint2D
     */
    inline size_t kdTreeClosestPoint3D(
        float x0, float y0, float z0, float& out_x, float& out_y, float& out_z,
        float& out_dist_sqr) const
    {
        MRPT_START
        rebuild_kdTree_3D();  // First: Create the 3D KD-Tree if required
        if (!m_kdtree3d_data.m_num_points)
            THROW_EXCEPTION("There are no points in the KD-tree.");

        const size_t knn = 1;  // Number of points to retrieve
        size_t ret_index;
        nanoflann::KNNResultSet<num_t> resultSet(knn);
        resultSet.init(&ret_index, &out_dist_sqr);

        const std::array<num_t, 3> query_point{{x0, y0, z0}};
        m_kdtree3d_data.index->findNeighbors(
            resultSet, &query_point[0], nanoflann::SearchParams());

        // Copy output to user vars:
        out_x = derived().kdtree_get_pt(ret_index, 0);
        out_y = derived().kdtree_get_pt(ret_index, 1);
        out_z = derived().kdtree_get_pt(ret_index, 2);

        return ret_index;
        MRPT_END
    }

    /// \overload
    inline size_t kdTreeClosestPoint3D(
        float x0, float y0, float z0, float& out_dist_sqr) const
    {
        MRPT_START
        rebuild_kdTree_3D();  // First: Create the 3D KD-Tree if required
        if (!m_kdtree3d_data.m_num_points)
            THROW_EXCEPTION("There are no points in the KD-tree.");

        const size_t knn = 1;  // Number of points to retrieve
        size_t ret_index;
        nanoflann::KNNResultSet<num_t> resultSet(knn);
        resultSet.init(&ret_index, &out_dist_sqr);

        const std::array<num_t, 3> query_point{{x0, y0, z0}};
        m_kdtree3d_data.index->findNeighbors(
            resultSet, &query_point[0], nanoflann::SearchParams());

        return ret_index;
        MRPT_END
    }

    /// \overload
    inline size_t kdTreeClosestPoint3D(
        const TPoint3D& p0, TPoint3D& pOut, float& outDistSqr) const
    {
        float dmy1, dmy2, dmy3;
        size_t res = kdTreeClosestPoint3D(
            static_cast<float>(p0.x), static_cast<float>(p0.y),
            static_cast<float>(p0.z), dmy1, dmy2, dmy3, outDistSqr);
        pOut.x = static_cast<double>(dmy1);
        pOut.y = static_cast<double>(dmy2);
        pOut.z = static_cast<double>(dmy3);
        return res;
    }

    /** KD Tree-based search for the N closest points to some given 3D
     *coordinates.
     *  This method automatically build the "m_kdtree_data" structure when:
     *      - It is called for the first time
     *      - The map has changed
     *      - The KD-tree was build for 2D.
     *
     * \param x0  The X coordinate of the query.
     * \param y0  The Y coordinate of the query.
     * \param z0  The Z coordinate of the query.
     * \param N The number of closest points to search.
     * \param out_x The vector containing the X coordinates of the
     *correspondences.
     * \param out_y The vector containing the Y coordinates of the
     *correspondences.
     * \param out_z The vector containing the Z coordinates of the
     *correspondences.
     * \param out_dist_sqr The vector containing the square distance between
     *the query and the returned points.
     *
     *  \sa kdTreeNClosestPoint2D
     */
    inline void kdTreeNClosestPoint3D(
        float x0, float y0, float z0, size_t knn, std::vector<float>& out_x,
        std::vector<float>& out_y, std::vector<float>& out_z,
        std::vector<float>& out_dist_sqr) const
    {
        MRPT_START
        rebuild_kdTree_3D();  // First: Create the 3D KD-Tree if required
        if (!m_kdtree3d_data.m_num_points)
            THROW_EXCEPTION("There are no points in the KD-tree.");

        std::vector<size_t> ret_indexes(knn);
        out_x.resize(knn);
        out_y.resize(knn);
        out_z.resize(knn);
        out_dist_sqr.resize(knn);

        nanoflann::KNNResultSet<num_t> resultSet(knn);
        resultSet.init(&ret_indexes[0], &out_dist_sqr[0]);

        const std::array<num_t, 3> query_point{{x0, y0, z0}};
        m_kdtree3d_data.index->findNeighbors(
            resultSet, &query_point[0], nanoflann::SearchParams());

        for (size_t i = 0; i < knn; i++)
        {
            out_x[i] = derived().kdtree_get_pt(ret_indexes[i], 0);
            out_y[i] = derived().kdtree_get_pt(ret_indexes[i], 1);
            out_z[i] = derived().kdtree_get_pt(ret_indexes[i], 2);
        }
        MRPT_END
    }

    /** KD Tree-based search for the N closest points to some given 3D
     *coordinates.
     *  This method automatically build the "m_kdtree_data" structure when:
     *      - It is called for the first time
     *      - The map has changed
     *      - The KD-tree was build for 2D.
     *
     * \param x0  The X coordinate of the query.
     * \param y0  The Y coordinate of the query.
     * \param z0  The Z coordinate of the query.
     * \param N The number of closest points to search.
     * \param out_x The vector containing the X coordinates of the
     *correspondences.
     * \param out_y The vector containing the Y coordinates of the
     *correspondences.
     * \param out_z The vector containing the Z coordinates of the
     *correspondences.
     * \param out_idx The vector containing the indexes of the correspondences.
     * \param out_dist_sqr The vector containing the square distance between
     *the query and the returned points.
     *
     *  \sa kdTreeNClosestPoint2D
     */
    inline void kdTreeNClosestPoint3DWithIdx(
        float x0, float y0, float z0, size_t knn, std::vector<float>& out_x,
        std::vector<float>& out_y, std::vector<float>& out_z,
        std::vector<size_t>& out_idx, std::vector<float>& out_dist_sqr) const
    {
        MRPT_START
        rebuild_kdTree_3D();  // First: Create the 3D KD-Tree if required
        if (!m_kdtree3d_data.m_num_points)
            THROW_EXCEPTION("There are no points in the KD-tree.");

        out_x.resize(knn);
        out_y.resize(knn);
        out_z.resize(knn);
        out_idx.resize(knn);
        out_dist_sqr.resize(knn);

        nanoflann::KNNResultSet<num_t> resultSet(knn);
        resultSet.init(&out_idx[0], &out_dist_sqr[0]);

        const std::array<num_t, 3> query_point{{x0, y0, z0}};
        m_kdtree3d_data.index->findNeighbors(
            resultSet, &query_point[0], nanoflann::SearchParams());

        for (size_t i = 0; i < knn; i++)
        {
            out_x[i] = derived().kdtree_get_pt(out_idx[i], 0);
            out_y[i] = derived().kdtree_get_pt(out_idx[i], 1);
            out_z[i] = derived().kdtree_get_pt(out_idx[i], 2);
        }
        MRPT_END
    }

    inline void kdTreeNClosestPoint3D(
        const TPoint3D& p0, size_t N, std::vector<TPoint3D>& pOut,
        std::vector<float>& outDistSqr) const
    {
        std::vector<float> dmy1, dmy2, dmy3;
        kdTreeNClosestPoint3D(
            static_cast<float>(p0.x), static_cast<float>(p0.y),
            static_cast<float>(p0.z), N, dmy1, dmy2, dmy3, outDistSqr);
        pOut.resize(dmy1.size());
        for (size_t i = 0; i < dmy1.size(); i++)
        {
            pOut[i].x = static_cast<double>(dmy1[i]);
            pOut[i].y = static_cast<double>(dmy2[i]);
            pOut[i].z = static_cast<double>(dmy3[i]);
        }
    }

    /** KD Tree-based search for all the points within a given radius of some 3D
     *point.
     *  This method automatically build the "m_kdtree_data" structure when:
     *      - It is called for the first time
     *      - The map has changed
     *      - The KD-tree was build for 2D.
     *
     * \param x0  The X coordinate of the query.
     * \param y0  The Y coordinate of the query.
     * \param z0  The Z coordinate of the query.
     * \param maxRadiusSqr The square of the desired search radius.
     * \param out_indices_dist The output list, with pairs of indeces/squared
     *distances for the found correspondences.
     * \return Number of found points.
     *
     *  \sa kdTreeRadiusSearch2D, kdTreeNClosestPoint3DIdx
     */
    inline size_t kdTreeRadiusSearch3D(
        const num_t x0, const num_t y0, const num_t z0,
        const num_t maxRadiusSqr,
        std::vector<std::pair<size_t, num_t>>& out_indices_dist) const
    {
        MRPT_START
        rebuild_kdTree_3D();  // First: Create the 3D KD-Tree if required
        out_indices_dist.clear();
        if (m_kdtree3d_data.m_num_points != 0)
        {
            const num_t xyz[3] = {x0, y0, z0};
            m_kdtree3d_data.index->radiusSearch(
                &xyz[0], maxRadiusSqr, out_indices_dist,
                nanoflann::SearchParams());
        }
        return out_indices_dist.size();
        MRPT_END
    }

    /** KD Tree-based search for all the points within a given radius of some 2D
     *point.
     *  This method automatically build the "m_kdtree_data" structure when:
     *      - It is called for the first time
     *      - The map has changed
     *      - The KD-tree was build for 3D.
     *
     * \param x0  The X coordinate of the query.
     * \param y0  The Y coordinate of the query.
     * \param maxRadiusSqr The square of the desired search radius.
     * \param out_indices_dist The output list, with pairs of indeces/squared
     *distances for the found correspondences.
     * \return Number of found points.
     *
     *  \sa kdTreeRadiusSearch3D, kdTreeNClosestPoint2DIdx
     */
    inline size_t kdTreeRadiusSearch2D(
        const num_t x0, const num_t y0, const num_t maxRadiusSqr,
        std::vector<std::pair<size_t, num_t>>& out_indices_dist) const
    {
        MRPT_START
        rebuild_kdTree_2D();  // First: Create the 2D KD-Tree if required
        out_indices_dist.clear();
        if (m_kdtree2d_data.m_num_points != 0)
        {
            const num_t xyz[2] = {x0, y0};
            m_kdtree2d_data.index->radiusSearch(
                &xyz[0], maxRadiusSqr, out_indices_dist,
                nanoflann::SearchParams());
        }
        return out_indices_dist.size();
        MRPT_END
    }

    /** KD Tree-based search for the N closest point to some given 3D
     *coordinates and returns their indexes.
     *  This method automatically build the "m_kdtree_data" structure when:
     *      - It is called for the first time
     *      - The map has changed
     *      - The KD-tree was build for 2D.
     *
     * \param x0  The X coordinate of the query.
     * \param y0  The Y coordinate of the query.
     * \param z0  The Z coordinate of the query.
     * \param N The number of closest points to search.
     * \param out_idx The indexes of the found closest correspondence.
     * \param out_dist_sqr The square distance between the query and the
     *returned point.
     *
     *  \sa kdTreeClosestPoint2D,  kdTreeRadiusSearch3D
     */
    inline void kdTreeNClosestPoint3DIdx(
        float x0, float y0, float z0, size_t knn, std::vector<size_t>& out_idx,
        std::vector<float>& out_dist_sqr) const
    {
        MRPT_START
        rebuild_kdTree_3D();  // First: Create the 3D KD-Tree if required
        if (!m_kdtree3d_data.m_num_points)
            THROW_EXCEPTION("There are no points in the KD-tree.");

        out_idx.resize(knn);
        out_dist_sqr.resize(knn);
        nanoflann::KNNResultSet<num_t> resultSet(knn);
        resultSet.init(&out_idx[0], &out_dist_sqr[0]);

        const std::array<num_t, 3> query_point{{x0, y0, z0}};
        m_kdtree3d_data.index->findNeighbors(
            resultSet, &query_point[0], nanoflann::SearchParams());
        MRPT_END
    }

    inline void kdTreeNClosestPoint3DIdx(
        const TPoint3D& p0, size_t N, std::vector<size_t>& outIdx,
        std::vector<float>& outDistSqr) const
    {
        kdTreeNClosestPoint3DIdx(
            static_cast<float>(p0.x), static_cast<float>(p0.y),
            static_cast<float>(p0.z), N, outIdx, outDistSqr);
    }

    inline void kdTreeEnsureIndexBuilt3D() { rebuild_kdTree_3D(); }
    inline void kdTreeEnsureIndexBuilt2D() { rebuild_kdTree_2D(); }

    /* @} */

   protected:
    /** To be called by child classes when KD tree data changes. */
    inline void kdtree_mark_as_outdated() const
    {
        std::lock_guard<std::mutex> lck(m_kdtree_mtx);
        m_kdtree_is_uptodate = false;
    }

   private:
    /** Internal structure with the KD-tree representation (mainly used to avoid
     * copying pointers with the = operator) */
    template <int _DIM = -1>
    struct TKDTreeDataHolder
    {
        inline TKDTreeDataHolder() = default;
        /** Copy constructor: It actually does NOT copy the kd-tree, a new
         * object will be created if required!   */
        inline TKDTreeDataHolder(const TKDTreeDataHolder&) : TKDTreeDataHolder()
        {
        }
        /** Copy operator: It actually does NOT copy the kd-tree, a new object
         * will be created if required!  */
        inline TKDTreeDataHolder& operator=(const TKDTreeDataHolder& o) noexcept
        {
            if (&o != this) clear();
            return *this;
        }

        /** Free memory (if allocated)  */
        inline void clear() noexcept { index.reset(); }
        using kdtree_index_t =
            nanoflann::KDTreeSingleIndexAdaptor<metric_t, Derived, _DIM>;

        /** nullptr or the up-to-date index */
        std::unique_ptr<kdtree_index_t> index;

        /** Dimensionality. typ: 2,3 */
        size_t m_dim = _DIM;
        size_t m_num_points = 0;
    };

    mutable std::mutex m_kdtree_mtx;
    mutable TKDTreeDataHolder<2> m_kdtree2d_data;
    mutable TKDTreeDataHolder<3> m_kdtree3d_data;
    /** whether the KD tree needs to be rebuilt or not. */
    mutable std::atomic_bool m_kdtree_is_uptodate{false};

    /// Rebuild, if needed the KD-tree for 2D (nDims=2), 3D (nDims=3), ...
    /// asking the child class for the data points.
    void rebuild_kdTree_2D() const
    {
        if (m_kdtree_is_uptodate) return;

        std::lock_guard<std::mutex> lck(m_kdtree_mtx);
        using tree2d_t = typename TKDTreeDataHolder<2>::kdtree_index_t;

        if (!m_kdtree_is_uptodate)
        {
            m_kdtree2d_data.clear();
            m_kdtree3d_data.clear();
        }

        if (!m_kdtree2d_data.index)
        {
            // Erase previous tree:
            m_kdtree2d_data.clear();
            // And build new index:
            const size_t N = derived().kdtree_get_point_count();
            m_kdtree2d_data.m_num_points = N;
            m_kdtree2d_data.m_dim = 2;
            if (N)
            {
                m_kdtree2d_data.index.reset(new tree2d_t(
                    2, derived(),
                    nanoflann::KDTreeSingleIndexAdaptorParams(
                        kdtree_search_params.leaf_max_size)));
                m_kdtree2d_data.index->buildIndex();
            }
            m_kdtree_is_uptodate = true;
        }
    }

    /// Rebuild, if needed the KD-tree for 2D (nDims=2), 3D (nDims=3), ...
    /// asking the child class for the data points.
    void rebuild_kdTree_3D() const
    {
        if (m_kdtree_is_uptodate) return;
        std::lock_guard<std::mutex> lck(m_kdtree_mtx);
        using tree3d_t = typename TKDTreeDataHolder<3>::kdtree_index_t;

        if (!m_kdtree_is_uptodate)
        {
            m_kdtree2d_data.clear();
            m_kdtree3d_data.clear();
        }

        if (!m_kdtree3d_data.index)
        {
            // Erase previous tree:
            m_kdtree3d_data.clear();
            // And build new index:
            const size_t N = derived().kdtree_get_point_count();
            m_kdtree3d_data.m_num_points = N;
            m_kdtree3d_data.m_dim = 3;
            if (N)
            {
                m_kdtree3d_data.index.reset(new tree3d_t(
                    3, derived(),
                    nanoflann::KDTreeSingleIndexAdaptorParams(
                        kdtree_search_params.leaf_max_size)));
                m_kdtree3d_data.index->buildIndex();
            }
            m_kdtree_is_uptodate = true;
        }
    }

};  // end of KDTreeCapable

/**  @} */  // end of grouping

}  // namespace mrpt::math