supervoxel_clustering.h

 
/*
 * Software License Agreement (BSD License)
 *
 *  Point Cloud Library (PCL) - www.pointclouds.org
 *
 *  All rights reserved.
 *
 *  Redistribution and use in source and binary forms, with or without
 *  modification, are permitted provided that the following conditions
 *  are met:
 *
 *   * Redistributions of source code must retain the above copyright
 *     notice, this list of conditions and the following disclaimer.
 *   * Redistributions in binary form must reproduce the above
 *     copyright notice, this list of conditions and the following
 *     disclaimer in the documentation and/or other materials provided
 *     with the distribution.
 *   * Neither the name of Willow Garage, Inc. nor the names of its
 *     contributors may be used to endorse or promote products derived
 *     from this software without specific prior written permission.
 *
 *  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 *  "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 *  LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
 *  FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
 *  COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
 *  INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
 *  BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 *  LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
 *  CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 *  LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
 *  ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 *  POSSIBILITY OF SUCH DAMAGE.
 *
 * Author : jpapon@gmail.com
 * Email  : jpapon@gmail.com
 *
 */
 
#pragma once
 
#include <boost/version.hpp>
 
#include <pcl/features/normal_3d.h>
#include <pcl/memory.h>
#include <pcl/pcl_base.h>
#include <pcl/pcl_macros.h>
#include <pcl/point_cloud.h>
#include <pcl/point_types.h>
#include <pcl/octree/octree_search.h>
#include <pcl/octree/octree_pointcloud_adjacency.h>
#include <pcl/search/search.h>
#include <boost/ptr_container/ptr_list.hpp> // for ptr_list
 
 
 
//DEBUG TODO REMOVE
#include <pcl/common/time.h>
 
 
namespace pcl
{
  /** \brief Supervoxel container class - stores a cluster extracted using supervoxel clustering
   */
  template <typename PointT>
  class Supervoxel
  {
    public:
      Supervoxel () :
        voxels_ (new pcl::PointCloud<PointT> ()),
        normals_ (new pcl::PointCloud<Normal> ())
        {  }
 
      using Ptr = shared_ptr<Supervoxel<PointT> >;
      using ConstPtr = shared_ptr<const Supervoxel<PointT> >;
 
      /** \brief Gets the centroid of the supervoxel
       *  \param[out] centroid_arg centroid of the supervoxel
       */
      void
      getCentroidPoint (PointXYZRGBA &centroid_arg)
      {
        centroid_arg = centroid_;
      }
 
      /** \brief Gets the point normal for the supervoxel
       * \param[out] normal_arg Point normal of the supervoxel
       * \note This isn't an average, it is a normal computed using all of the voxels in the supervoxel as support
       */
      void
      getCentroidPointNormal (PointNormal &normal_arg)
      {
        normal_arg.x = centroid_.x;
        normal_arg.y = centroid_.y;
        normal_arg.z = centroid_.z;
        normal_arg.normal_x = normal_.normal_x;
        normal_arg.normal_y = normal_.normal_y;
        normal_arg.normal_z = normal_.normal_z;
        normal_arg.curvature = normal_.curvature;
      }
 
      /** \brief The normal calculated for the voxels contained in the supervoxel */
      pcl::Normal normal_;
      /** \brief The centroid of the supervoxel - average voxel */
      pcl::PointXYZRGBA centroid_;
      /** \brief A Pointcloud of the voxels in the supervoxel */
      typename pcl::PointCloud<PointT>::Ptr voxels_;
      /** \brief A Pointcloud of the normals for the points in the supervoxel */
      typename pcl::PointCloud<Normal>::Ptr normals_;
 
    public:
      PCL_MAKE_ALIGNED_OPERATOR_NEW
  };
 
  /** \brief Implements a supervoxel algorithm based on voxel structure, normals, and rgb values
   *   \note Supervoxels are oversegmented volumetric patches (usually surfaces)
   *   \note Usually, color isn't needed (and can be detrimental)- spatial structure is mainly used
    * - J. Papon, A. Abramov, M. Schoeler, F. Woergoetter
    *   Voxel Cloud Connectivity Segmentation - Supervoxels from PointClouds
    *   In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2013
    *  \ingroup segmentation
    *  \author Jeremie Papon (jpapon@gmail.com)
    */
  template <typename PointT>
  class PCL_EXPORTS SupervoxelClustering : public pcl::PCLBase<PointT>
  {
    //Forward declaration of friended helper class
    class SupervoxelHelper;
    friend class SupervoxelHelper;
    public:
      /** \brief VoxelData is a structure used for storing data within a pcl::octree::OctreePointCloudAdjacencyContainer
       *  \note It stores xyz, rgb, normal, distance, an index, and an owner.
       */
      class VoxelData
      {
        public:
          VoxelData ():
            xyz_ (0.0f, 0.0f, 0.0f),
            rgb_ (0.0f, 0.0f, 0.0f),
            normal_ (0.0f, 0.0f, 0.0f, 0.0f),
            curvature_ (0.0f),
            distance_(0),
            idx_(0),
            owner_ (nullptr)
            {}
 
          /** \brief Gets the data of in the form of a point
           *  \param[out] point_arg Will contain the point value of the voxeldata
           */
          void
          getPoint (PointT &point_arg) const;
 
          /** \brief Gets the data of in the form of a normal
           *  \param[out] normal_arg Will contain the normal value of the voxeldata
           */
          void
          getNormal (Normal &normal_arg) const;
 
          Eigen::Vector3f xyz_;
          Eigen::Vector3f rgb_;
          Eigen::Vector4f normal_;
          float curvature_;
          float distance_;
          int idx_;
          SupervoxelHelper* owner_;
 
        public:
          PCL_MAKE_ALIGNED_OPERATOR_NEW
      };
 
      using LeafContainerT = pcl::octree::OctreePointCloudAdjacencyContainer<PointT, VoxelData>;
      using LeafVectorT = std::vector<LeafContainerT *>;
 
      using PointCloudT = pcl::PointCloud<PointT>;
      using NormalCloudT = pcl::PointCloud<Normal>;
      using OctreeAdjacencyT = pcl::octree::OctreePointCloudAdjacency<PointT, LeafContainerT>;
      using OctreeSearchT = pcl::octree::OctreePointCloudSearch<PointT>;
      using KdTreeT = pcl::search::KdTree<PointT>;
      using IndicesPtr = pcl::IndicesPtr;
 
      using PCLBase <PointT>::initCompute;
      using PCLBase <PointT>::deinitCompute;
      using PCLBase <PointT>::input_;
 
      using VoxelAdjacencyList = boost::adjacency_list<boost::setS, boost::setS, boost::undirectedS, std::uint32_t, float>;
      using VoxelID = VoxelAdjacencyList::vertex_descriptor;
      using EdgeID = VoxelAdjacencyList::edge_descriptor;
 
    public:
 
      /** \brief Constructor that sets default values for member variables.
       *  \param[in] voxel_resolution The resolution (in meters) of voxels used
       *  \param[in] seed_resolution The average size (in meters) of resulting supervoxels
       */
      SupervoxelClustering (float voxel_resolution, float seed_resolution);
 
      /** \brief This destructor destroys the cloud, normals and search method used for
        * finding neighbors. In other words it frees memory.
        */
 
      ~SupervoxelClustering () override;
 
      /** \brief Set the resolution of the octree voxels */
      void
      setVoxelResolution (float resolution);
 
      /** \brief Get the resolution of the octree voxels */
      float
      getVoxelResolution () const;
 
      /** \brief Set the resolution of the octree seed voxels */
      void
      setSeedResolution (float seed_resolution);
 
      /** \brief Get the resolution of the octree seed voxels */
      float
      getSeedResolution () const;
 
      /** \brief Set the importance of color for supervoxels */
      void
      setColorImportance (float val);
 
      /** \brief Set the importance of spatial distance for supervoxels */
      void
      setSpatialImportance (float val);
 
      /** \brief Set the importance of scalar normal product for supervoxels */
      void
      setNormalImportance (float val);
 
      /** \brief Set whether or not to use the single camera transform
       *  \note By default it will be used for organized clouds, but not for unorganized - this parameter will override that behavior
       *  The single camera transform scales bin size so that it increases exponentially with depth (z dimension).
       *  This is done to account for the decreasing point density found with depth when using an RGB-D camera.
       *  Without the transform, beyond a certain depth adjacency of voxels breaks down unless the voxel size is set to a large value.
       *  Using the transform allows preserving detail up close, while allowing adjacency at distance.
       *  The specific transform used here is:
       *  x /= z; y /= z; z = ln(z);
       *  This transform is applied when calculating the octree bins in OctreePointCloudAdjacency
       */
      void
      setUseSingleCameraTransform (bool val);
 
      /** \brief This method launches the segmentation algorithm and returns the supervoxels that were
       * obtained during the segmentation.
       * \param[out] supervoxel_clusters A map of labels to pointers to supervoxel structures
       */
      virtual void
      extract (std::map<std::uint32_t,typename Supervoxel<PointT>::Ptr > &supervoxel_clusters);
 
      /** \brief This method sets the cloud to be supervoxelized
       * \param[in] cloud The cloud to be supervoxelize
       */
      void
      setInputCloud (const typename pcl::PointCloud<PointT>::ConstPtr& cloud) override;
 
      /** \brief This method sets the normals to be used for supervoxels (should be same size as input cloud)
      * \param[in] normal_cloud The input normals
      */
      virtual void
      setNormalCloud (typename NormalCloudT::ConstPtr normal_cloud);
 
      /** \brief This method refines the calculated supervoxels - may only be called after extract
       * \param[in] num_itr The number of iterations of refinement to be done (2 or 3 is usually sufficient)
       * \param[out] supervoxel_clusters The resulting refined supervoxels
       */
      virtual void
      refineSupervoxels (int num_itr, std::map<std::uint32_t,typename Supervoxel<PointT>::Ptr > &supervoxel_clusters);
 
      ////////////////////////////////////////////////////////////
      /** \brief Returns a deep copy of the voxel centroid cloud */
      typename pcl::PointCloud<PointT>::Ptr
      getVoxelCentroidCloud () const;
 
      /** \brief Returns labeled cloud
        * Points that belong to the same supervoxel have the same label.
        * Labels for segments start from 1, unlabeled points have label 0
        */
      typename pcl::PointCloud<PointXYZL>::Ptr
      getLabeledCloud () const;
 
      /** \brief Returns labeled voxelized cloud
       * Points that belong to the same supervoxel have the same label.
       * Labels for segments start from 1, unlabeled points have label 0
       */
      pcl::PointCloud<pcl::PointXYZL>::Ptr
      getLabeledVoxelCloud () const;
 
      /** \brief Gets the adjacency list (Boost Graph library) which gives connections between supervoxels
       *  \param[out] adjacency_list_arg BGL graph where supervoxel labels are vertices, edges are touching relationships
       */
      void
      getSupervoxelAdjacencyList (VoxelAdjacencyList &adjacency_list_arg) const;
 
      /** \brief Get a multimap which gives supervoxel adjacency
       *  \param[out] label_adjacency Multi-Map which maps a supervoxel label to all adjacent supervoxel labels
       */
      void
      getSupervoxelAdjacency (std::multimap<std::uint32_t, std::uint32_t> &label_adjacency) const;
 
      /** \brief Static helper function which returns a pointcloud of normals for the input supervoxels
       *  \param[in] supervoxel_clusters Supervoxel cluster map coming from this class
       *  \returns Cloud of PointNormals of the supervoxels
       *
       */
      static pcl::PointCloud<pcl::PointNormal>::Ptr
      makeSupervoxelNormalCloud (std::map<std::uint32_t,typename Supervoxel<PointT>::Ptr > &supervoxel_clusters);
 
      /** \brief Returns the current maximum (highest) label */
      int
      getMaxLabel () const;
 
    private:
      /** \brief This method simply checks if it is possible to execute the segmentation algorithm with
        * the current settings. If it is possible then it returns true.
        */
      virtual bool
      prepareForSegmentation ();
 
      /** \brief This selects points to use as initial supervoxel centroids
       *  \param[out] seed_indices The selected leaf indices
       */
      void
      selectInitialSupervoxelSeeds (Indices &seed_indices);
 
      /** \brief This method creates the internal supervoxel helpers based on the provided seed points
       *  \param[in] seed_indices Indices of the leaves to use as seeds
       */
      void
      createSupervoxelHelpers (Indices &seed_indices);
 
      /** \brief This performs the superpixel evolution */
      void
      expandSupervoxels (int depth);
 
      /** \brief This sets the data of the voxels in the tree */
      void
      computeVoxelData ();
 
      /** \brief Reseeds the supervoxels by finding the voxel closest to current centroid */
      void
      reseedSupervoxels ();
 
      /** \brief Constructs the map of supervoxel clusters from the internal supervoxel helpers */
      void
      makeSupervoxels (std::map<std::uint32_t,typename Supervoxel<PointT>::Ptr > &supervoxel_clusters);
 
      /** \brief Stores the resolution used in the octree */
      float resolution_;
 
      /** \brief Stores the resolution used to seed the superpixels */
      float seed_resolution_;
 
      /** \brief Distance function used for comparing voxelDatas */
      float
      voxelDataDistance (const VoxelData &v1, const VoxelData &v2) const;
 
      /** \brief Transform function used to normalize voxel density versus distance from camera */
      void
      transformFunction (PointT &p);
 
      /** \brief Contains a KDtree for the voxelized cloud */
      typename pcl::search::KdTree<PointT>::Ptr voxel_kdtree_;
 
      /** \brief Octree Adjacency structure with leaves at voxel resolution */
      typename OctreeAdjacencyT::Ptr adjacency_octree_;
 
      /** \brief Contains the Voxelized centroid Cloud */
      typename PointCloudT::Ptr voxel_centroid_cloud_;
 
      /** \brief Contains the Voxelized centroid Cloud */
      typename NormalCloudT::ConstPtr input_normals_;
 
      /** \brief Importance of color in clustering */
      float color_importance_;
      /** \brief Importance of distance from seed center in clustering */
      float spatial_importance_;
      /** \brief Importance of similarity in normals for clustering */
      float normal_importance_;
 
      /** \brief Whether or not to use the transform compressing depth in Z
       *  This is only checked if it has been manually set by the user.
       *  The default behavior is to use the transform for organized, and not for unorganized.
       */
      bool use_single_camera_transform_;
      /** \brief Whether to use default transform behavior or not */
      bool use_default_transform_behaviour_;
 
      /** \brief Internal storage class for supervoxels
       * \note Stores pointers to leaves of clustering internal octree,
       * \note so should not be used outside of clustering class
       */
      class SupervoxelHelper
      {
        public:
          /** \brief Comparator for LeafContainerT pointers - used for sorting set of leaves
           * \note Compares by index in the overall leaf_vector. Order isn't important, so long as it is fixed.
           */
          struct compareLeaves
          {
            bool operator() (LeafContainerT* const &left, LeafContainerT* const &right) const
            {
              const VoxelData& leaf_data_left = left->getData ();
              const VoxelData& leaf_data_right = right->getData ();
              return leaf_data_left.idx_ < leaf_data_right.idx_;
            }
          };
          using LeafSetT = std::set<LeafContainerT*, typename SupervoxelHelper::compareLeaves>;
          using iterator = typename LeafSetT::iterator;
          using const_iterator = typename LeafSetT::const_iterator;
 
          SupervoxelHelper (std::uint32_t label, SupervoxelClustering* parent_arg):
            label_ (label),
            parent_ (parent_arg)
          { }
 
          void
          addLeaf (LeafContainerT* leaf_arg);
 
          void
          removeLeaf (LeafContainerT* leaf_arg);
 
          void
          removeAllLeaves ();
 
          void
          expand ();
 
          void
          refineNormals ();
 
          void
          updateCentroid ();
 
          void
          getVoxels (typename pcl::PointCloud<PointT>::Ptr &voxels) const;
 
          void
          getNormals (typename pcl::PointCloud<Normal>::Ptr &normals) const;
 
          using DistFuncPtr = float (SupervoxelClustering<PointT>::*)(const VoxelData &, const VoxelData &);
 
          std::uint32_t
          getLabel () const
          { return label_; }
 
          Eigen::Vector4f
          getNormal () const
          { return centroid_.normal_; }
 
          Eigen::Vector3f
          getRGB () const
          { return centroid_.rgb_; }
 
          Eigen::Vector3f
          getXYZ () const
          { return centroid_.xyz_;}
 
          void
          getXYZ (float &x, float &y, float &z) const
          { x=centroid_.xyz_[0]; y=centroid_.xyz_[1]; z=centroid_.xyz_[2]; }
 
          void
          getRGB (std::uint32_t &rgba) const
          {
            rgba = static_cast<std::uint32_t>(centroid_.rgb_[0]) << 16 |
                   static_cast<std::uint32_t>(centroid_.rgb_[1]) << 8 |
                   static_cast<std::uint32_t>(centroid_.rgb_[2]);
          }
 
          void
          getNormal (pcl::Normal &normal_arg) const
          {
            normal_arg.normal_x = centroid_.normal_[0];
            normal_arg.normal_y = centroid_.normal_[1];
            normal_arg.normal_z = centroid_.normal_[2];
            normal_arg.curvature = centroid_.curvature_;
          }
 
          void
          getNeighborLabels (std::set<std::uint32_t> &neighbor_labels) const;
 
          VoxelData
          getCentroid () const
          { return centroid_; }
 
          std::size_t
          size () const { return leaves_.size (); }
        private:
          //Stores leaves
          LeafSetT leaves_;
          std::uint32_t label_;
          VoxelData centroid_;
          SupervoxelClustering* parent_;
        public:
          //Type VoxelData may have fixed-size Eigen objects inside
          PCL_MAKE_ALIGNED_OPERATOR_NEW
      };
 
      //Make boost::ptr_list can access the private class SupervoxelHelper
#if BOOST_VERSION >= 107000
      friend void boost::checked_delete<> (const typename pcl::SupervoxelClustering<PointT>::SupervoxelHelper *) BOOST_NOEXCEPT;
#else
      friend void boost::checked_delete<> (const typename pcl::SupervoxelClustering<PointT>::SupervoxelHelper *);
#endif
 
      using HelperListT = boost::ptr_list<SupervoxelHelper>;
      HelperListT supervoxel_helpers_;
 
      //TODO DEBUG REMOVE
      StopWatch timer_;
    public:
      PCL_MAKE_ALIGNED_OPERATOR_NEW
  };
 
}
 
#ifdef PCL_NO_PRECOMPILE
#include <pcl/segmentation/impl/supervoxel_clustering.hpp>
#endif