supervoxel_clustering.hpp

/*
 * 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
 *
 */
 
#ifndef PCL_SEGMENTATION_SUPERVOXEL_CLUSTERING_HPP_
#define PCL_SEGMENTATION_SUPERVOXEL_CLUSTERING_HPP_
 
#include <pcl/segmentation/supervoxel_clustering.h>
#include <pcl/common/io.h> // for copyPointCloud
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT>
pcl::SupervoxelClustering<PointT>::SupervoxelClustering (float voxel_resolution, float seed_resolution) :
  resolution_ (voxel_resolution),
  seed_resolution_ (seed_resolution),
  adjacency_octree_ (),
  voxel_centroid_cloud_ (),
  color_importance_ (0.1f),
  spatial_importance_ (0.4f),
  normal_importance_ (1.0f),
  use_default_transform_behaviour_ (true)
{
  adjacency_octree_.reset (new OctreeAdjacencyT (resolution_));
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT>
pcl::SupervoxelClustering<PointT>::~SupervoxelClustering () = default;
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::setInputCloud (const typename pcl::PointCloud<PointT>::ConstPtr& cloud)
{
  if ( cloud->empty () )
  {
    PCL_ERROR ("[pcl::SupervoxelClustering::setInputCloud] Empty cloud set, doing nothing \n");
    return;
  }
  
  input_ = cloud;
  adjacency_octree_->setInputCloud (cloud);
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::setNormalCloud (typename NormalCloudT::ConstPtr normal_cloud)
{
  if ( normal_cloud->empty () )
  {
    PCL_ERROR ("[pcl::SupervoxelClustering::setNormalCloud] Empty cloud set, doing nothing \n");
    return;
  }
  
  input_normals_ = normal_cloud;
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::extract (std::map<std::uint32_t,typename Supervoxel<PointT>::Ptr > &supervoxel_clusters)
{
  //timer_.reset ();
  //double t_start = timer_.getTime ();
  //std::cout << "Init compute  \n";
  bool segmentation_is_possible = initCompute ();
  if ( !segmentation_is_possible )
  {
    deinitCompute ();
    return;
  }
  
  //std::cout << "Preparing for segmentation \n";
  segmentation_is_possible = prepareForSegmentation ();
  if ( !segmentation_is_possible )
  {
    deinitCompute ();
    return;
  }
  
  //double t_prep = timer_.getTime ();
  //std::cout << "Placing Seeds" << std::endl;
  Indices seed_indices;
  selectInitialSupervoxelSeeds (seed_indices);
  //std::cout << "Creating helpers "<<std::endl;
  createSupervoxelHelpers (seed_indices);
  //double t_seeds = timer_.getTime ();
  
  
  //std::cout << "Expanding the supervoxels" << std::endl;
  int max_depth = static_cast<int> (1.8f*seed_resolution_/resolution_);
  expandSupervoxels (max_depth);
  //double t_iterate = timer_.getTime ();
    
  //std::cout << "Making Supervoxel structures" << std::endl;
  makeSupervoxels (supervoxel_clusters);
  //double t_supervoxels = timer_.getTime ();
  
 // std::cout << "--------------------------------- Timing Report --------------------------------- \n";
 // std::cout << "Time to prep (normals, neighbors, voxelization)="<<t_prep-t_start<<" ms\n";
 // std::cout << "Time to seed clusters                          ="<<t_seeds-t_prep<<" ms\n";
 // std::cout << "Time to expand clusters                        ="<<t_iterate-t_seeds<<" ms\n";
 // std::cout << "Time to create supervoxel structures           ="<<t_supervoxels-t_iterate<<" ms\n";
 // std::cout << "Total run time                                 ="<<t_supervoxels-t_start<<" ms\n";
 // std::cout << "--------------------------------------------------------------------------------- \n";
  
  deinitCompute ();
}
 
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::refineSupervoxels (int num_itr, std::map<std::uint32_t,typename Supervoxel<PointT>::Ptr > &supervoxel_clusters)
{
  if (supervoxel_helpers_.empty ())
  {
    PCL_ERROR ("[pcl::SupervoxelClustering::refineVoxelNormals] Supervoxels not extracted, doing nothing - (Call extract first!) \n");
    return;
  }
 
  int max_depth = static_cast<int> (1.8f*seed_resolution_/resolution_);
  for (int i = 0; i < num_itr; ++i)
  {
    for (typename HelperListT::iterator sv_itr = supervoxel_helpers_.begin (); sv_itr != supervoxel_helpers_.end (); ++sv_itr)
    {
      sv_itr->refineNormals ();
    }
    
    reseedSupervoxels ();
    expandSupervoxels (max_depth);
  }
  
 
  makeSupervoxels (supervoxel_clusters);
 
}
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
 
template <typename PointT> bool
pcl::SupervoxelClustering<PointT>::prepareForSegmentation ()
{
  
  // if user forgot to pass point cloud or if it is empty
  if ( input_->points.empty () )
    return (false);
  
  //Add the new cloud of data to the octree
  //std::cout << "Populating adjacency octree with new cloud \n";
  //double prep_start = timer_.getTime ();
  if ( (use_default_transform_behaviour_ && input_->isOrganized ())
       || (!use_default_transform_behaviour_ && use_single_camera_transform_))
      adjacency_octree_->setTransformFunction ([this] (PointT &p) { transformFunction (p); });
 
  adjacency_octree_->addPointsFromInputCloud ();
  //double prep_end = timer_.getTime ();
  //std::cout<<"Time elapsed populating octree with next frame ="<<prep_end-prep_start<<" ms\n";
  
  //Compute normals and insert data for centroids into data field of octree
  //double normals_start = timer_.getTime ();
  computeVoxelData ();
  //double normals_end = timer_.getTime ();
  //std::cout << "Time elapsed finding normals and pushing into octree ="<<normals_end-normals_start<<" ms\n";
    
  return true;
}
 
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::computeVoxelData ()
{
  voxel_centroid_cloud_.reset (new PointCloudT);
  voxel_centroid_cloud_->resize (adjacency_octree_->getLeafCount ());
  auto leaf_itr = adjacency_octree_->begin ();
  auto cent_cloud_itr = voxel_centroid_cloud_->begin ();
  for (int idx = 0 ; leaf_itr != adjacency_octree_->end (); ++leaf_itr, ++cent_cloud_itr, ++idx)
  {
    VoxelData& new_voxel_data = (*leaf_itr)->getData ();
    //Add the point to the centroid cloud
    new_voxel_data.getPoint (*cent_cloud_itr);
    //voxel_centroid_cloud_->push_back(new_voxel_data.getPoint ());
    new_voxel_data.idx_ = idx;
  }
  
  //If normals were provided
  if (input_normals_)
  {
    //Verify that input normal cloud size is same as input cloud size
    assert (input_normals_->size () == input_->size ());
    //For every point in the input cloud, find its corresponding leaf
    auto normal_itr = input_normals_->begin ();
    for (auto input_itr = input_->begin (); input_itr != input_->end (); ++input_itr, ++normal_itr)
    {
      //If the point is not finite we ignore it
      if ( !pcl::isFinite<PointT> (*input_itr))
        continue;
      //Otherwise look up its leaf container
      LeafContainerT* leaf = adjacency_octree_->getLeafContainerAtPoint (*input_itr);
 
      //Get the voxel data object
      VoxelData& voxel_data = leaf->getData ();
      //Add this normal in (we will normalize at the end)
      voxel_data.normal_ += normal_itr->getNormalVector4fMap ();
      voxel_data.curvature_ += normal_itr->curvature;
    }
    //Now iterate through the leaves and normalize 
    for (leaf_itr = adjacency_octree_->begin (); leaf_itr != adjacency_octree_->end (); ++leaf_itr)
    {
      VoxelData& voxel_data = (*leaf_itr)->getData ();
      voxel_data.normal_.normalize ();
      voxel_data.owner_ = nullptr;
      voxel_data.distance_ = std::numeric_limits<float>::max ();
      //Get the number of points in this leaf
      int num_points = (*leaf_itr)->getPointCounter ();
      voxel_data.curvature_ /= num_points;
    }
  }
  else //Otherwise just compute the normals
  {
    for (leaf_itr = adjacency_octree_->begin (); leaf_itr != adjacency_octree_->end (); ++leaf_itr)
    {
      VoxelData& new_voxel_data = (*leaf_itr)->getData ();
      //For every point, get its neighbors, build an index vector, compute normal
      Indices indices;
      indices.reserve (81); 
      //Push this point
      indices.push_back (new_voxel_data.idx_);
      for (auto neighb_itr=(*leaf_itr)->cbegin (); neighb_itr!=(*leaf_itr)->cend (); ++neighb_itr)
      {
        VoxelData& neighb_voxel_data = (*neighb_itr)->getData ();
        //Push neighbor index
        indices.push_back (neighb_voxel_data.idx_);
        //Get neighbors neighbors, push onto cloud
        for (auto neighb_neighb_itr=(*neighb_itr)->cbegin (); neighb_neighb_itr!=(*neighb_itr)->cend (); ++neighb_neighb_itr)
        {
          VoxelData& neighb2_voxel_data = (*neighb_neighb_itr)->getData ();
          indices.push_back (neighb2_voxel_data.idx_);
        }
      }
      //Compute normal
      pcl::computePointNormal (*voxel_centroid_cloud_, indices, new_voxel_data.normal_, new_voxel_data.curvature_);
      pcl::flipNormalTowardsViewpoint ((*voxel_centroid_cloud_)[new_voxel_data.idx_], 0.0f,0.0f,0.0f, new_voxel_data.normal_);
      new_voxel_data.normal_[3] = 0.0f;
      new_voxel_data.normal_.normalize ();
      new_voxel_data.owner_ = nullptr;
      new_voxel_data.distance_ = std::numeric_limits<float>::max ();
    }
  }
  
  
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::expandSupervoxels ( int depth )
{
  
  
  for (int i = 1; i < depth; ++i)
  {
      //Expand the the supervoxels by one iteration
      for (typename HelperListT::iterator sv_itr = supervoxel_helpers_.begin (); sv_itr != supervoxel_helpers_.end (); ++sv_itr)
      {
        sv_itr->expand ();
      }
      
      //Update the centers to reflect new centers
      for (typename HelperListT::iterator sv_itr = supervoxel_helpers_.begin (); sv_itr != supervoxel_helpers_.end (); )
      {
        if (sv_itr->size () == 0)
        {
          sv_itr = supervoxel_helpers_.erase (sv_itr);
        }
        else
        {
          sv_itr->updateCentroid ();
          ++sv_itr;
        } 
      }
 
  }
 
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::makeSupervoxels (std::map<std::uint32_t,typename Supervoxel<PointT>::Ptr > &supervoxel_clusters)
{
  supervoxel_clusters.clear ();
  for (typename HelperListT::iterator sv_itr = supervoxel_helpers_.begin (); sv_itr != supervoxel_helpers_.end (); ++sv_itr)
  {
    std::uint32_t label = sv_itr->getLabel ();
    supervoxel_clusters[label].reset (new Supervoxel<PointT>);
    sv_itr->getXYZ (supervoxel_clusters[label]->centroid_.x,supervoxel_clusters[label]->centroid_.y,supervoxel_clusters[label]->centroid_.z);
    sv_itr->getRGB (supervoxel_clusters[label]->centroid_.rgba);
    sv_itr->getNormal (supervoxel_clusters[label]->normal_);
    sv_itr->getVoxels (supervoxel_clusters[label]->voxels_);
    sv_itr->getNormals (supervoxel_clusters[label]->normals_);
  }
}
 
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::createSupervoxelHelpers (Indices &seed_indices)
{
  
  supervoxel_helpers_.clear ();
  for (std::size_t i = 0; i < seed_indices.size (); ++i)
  {
    supervoxel_helpers_.push_back (new SupervoxelHelper(i+1,this));
    //Find which leaf corresponds to this seed index
    LeafContainerT* seed_leaf = adjacency_octree_->at(seed_indices[i]);//adjacency_octree_->getLeafContainerAtPoint (seed_points[i]);
    if (seed_leaf)
    {
      supervoxel_helpers_.back ().addLeaf (seed_leaf);
    }
    else
    {
      PCL_WARN ("Could not find leaf in pcl::SupervoxelClustering<PointT>::createSupervoxelHelpers - supervoxel will be deleted \n");
    }
  }
  
}
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::selectInitialSupervoxelSeeds (Indices &seed_indices)
{
  //TODO THIS IS BAD - SEEDING SHOULD BE BETTER
  //TODO Switch to assigning leaves! Don't use Octree!
  
 // std::cout << "Size of centroid cloud="<<voxel_centroid_cloud_->size ()<<", seeding resolution="<<seed_resolution_<<"\n";
  //Initialize octree with voxel centroids
  pcl::octree::OctreePointCloudSearch <PointT> seed_octree (seed_resolution_);
  seed_octree.setInputCloud (voxel_centroid_cloud_);
  seed_octree.addPointsFromInputCloud ();
 // std::cout << "Size of octree ="<<seed_octree.getLeafCount ()<<"\n";
  std::vector<PointT, Eigen::aligned_allocator<PointT> > voxel_centers; 
  int num_seeds = seed_octree.getOccupiedVoxelCenters(voxel_centers); 
  //std::cout << "Number of seed points before filtering="<<voxel_centers.size ()<<std::endl;
  
  std::vector<int> seed_indices_orig;
  seed_indices_orig.resize (num_seeds, 0);
  seed_indices.clear ();
  pcl::Indices closest_index;
  std::vector<float> distance;
  closest_index.resize(1,0);
  distance.resize(1,0);
  if (!voxel_kdtree_)
  {
    voxel_kdtree_.reset (new pcl::search::KdTree<PointT>);
    voxel_kdtree_ ->setInputCloud (voxel_centroid_cloud_);
  }
  
  for (int i = 0; i < num_seeds; ++i)  
  {
    voxel_kdtree_->nearestKSearch (voxel_centers[i], 1, closest_index, distance);
    seed_indices_orig[i] = closest_index[0];
  }
  
  pcl::Indices neighbors;
  std::vector<float> sqr_distances;
  seed_indices.reserve (seed_indices_orig.size ());
  float search_radius = 0.5f*seed_resolution_;
  // This is 1/20th of the number of voxels which fit in a planar slice through search volume
  // Area of planar slice / area of voxel side. (Note: This is smaller than the value mentioned in the original paper)
  float min_points = 0.05f * (search_radius)*(search_radius) * 3.1415926536f  / (resolution_*resolution_);
  for (const auto &index_orig : seed_indices_orig)
  {
    int num = voxel_kdtree_->radiusSearch (index_orig, search_radius , neighbors, sqr_distances);
    if ( num > min_points)
    {
      seed_indices.push_back (index_orig);
    }
    
  }
 // std::cout << "Number of seed points after filtering="<<seed_points.size ()<<std::endl;
  
}
 
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::reseedSupervoxels ()
{
  //Go through each supervoxel and remove all it's leaves
  for (typename HelperListT::iterator sv_itr = supervoxel_helpers_.begin (); sv_itr != supervoxel_helpers_.end (); ++sv_itr)
  {
    sv_itr->removeAllLeaves ();
  }
  
  Indices closest_index;
  std::vector<float> distance;
  //Now go through each supervoxel, find voxel closest to its center, add it in
  for (typename HelperListT::iterator sv_itr = supervoxel_helpers_.begin (); sv_itr != supervoxel_helpers_.end (); ++sv_itr)
  {
    PointT point;
    sv_itr->getXYZ (point.x, point.y, point.z);
    voxel_kdtree_->nearestKSearch (point, 1, closest_index, distance);
    
    LeafContainerT* seed_leaf = adjacency_octree_->at (closest_index[0]);
    if (seed_leaf)
    {
      sv_itr->addLeaf (seed_leaf);
    }
    else
    {
      PCL_WARN ("Could not find leaf in pcl::SupervoxelClustering<PointT>::reseedSupervoxels - supervoxel will be deleted \n");
    }
  }
  
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::transformFunction (PointT &p)
{
  p.x /= p.z;
  p.y /= p.z;
  p.z = std::log (p.z);
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> float
pcl::SupervoxelClustering<PointT>::voxelDataDistance (const VoxelData &v1, const VoxelData &v2) const
{
  
  float spatial_dist = (v1.xyz_ - v2.xyz_).norm () / seed_resolution_;
  float color_dist =  (v1.rgb_ - v2.rgb_).norm () / 255.0f;
  float cos_angle_normal = 1.0f - std::abs (v1.normal_.dot (v2.normal_));
 // std::cout << "s="<<spatial_dist<<"  c="<<color_dist<<"   an="<<cos_angle_normal<<"\n";
  return  cos_angle_normal * normal_importance_ + color_dist * color_importance_+ spatial_dist * spatial_importance_;
  
}
 
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
///////// GETTER FUNCTIONS
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::getSupervoxelAdjacencyList (VoxelAdjacencyList &adjacency_list_arg) const 
{
  adjacency_list_arg.clear ();
    //Add a vertex for each label, store ids in map
  std::map <std::uint32_t, VoxelID> label_ID_map;
  for (typename HelperListT::const_iterator sv_itr = supervoxel_helpers_.cbegin (); sv_itr != supervoxel_helpers_.cend (); ++sv_itr)
  {
    VoxelID node_id = add_vertex (adjacency_list_arg);
    adjacency_list_arg[node_id] = (sv_itr->getLabel ());
    label_ID_map.insert (std::make_pair (sv_itr->getLabel (), node_id));
  }
  
  for (typename HelperListT::const_iterator sv_itr = supervoxel_helpers_.cbegin (); sv_itr != supervoxel_helpers_.cend (); ++sv_itr)
  {
    std::uint32_t label = sv_itr->getLabel ();
    std::set<std::uint32_t> neighbor_labels;
    sv_itr->getNeighborLabels (neighbor_labels);
    for (const unsigned int &neighbor_label : neighbor_labels)
    {
      bool edge_added;
      EdgeID edge;
      VoxelID u = (label_ID_map.find (label))->second;
      VoxelID v = (label_ID_map.find (neighbor_label))->second;
      boost::tie (edge, edge_added) = add_edge (u,v,adjacency_list_arg);
      //Calc distance between centers, set as edge weight
      if (edge_added)
      {
        VoxelData centroid_data = (sv_itr)->getCentroid ();
        //Find the neighbor with this label
        VoxelData neighb_centroid_data;
        
        for (typename HelperListT::const_iterator neighb_itr = supervoxel_helpers_.cbegin (); neighb_itr != supervoxel_helpers_.cend (); ++neighb_itr)
        {
          if (neighb_itr->getLabel () == neighbor_label)
          {
            neighb_centroid_data = neighb_itr->getCentroid ();
            break;
          }
        }
        
        float length = voxelDataDistance (centroid_data, neighb_centroid_data);
        adjacency_list_arg[edge] = length;
      }
    }
      
  }
 
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::getSupervoxelAdjacency (std::multimap<std::uint32_t, std::uint32_t> &label_adjacency) const
{
  label_adjacency.clear ();
  for (typename HelperListT::const_iterator sv_itr = supervoxel_helpers_.cbegin (); sv_itr != supervoxel_helpers_.cend (); ++sv_itr)
  {
    std::uint32_t label = sv_itr->getLabel ();
    std::set<std::uint32_t> neighbor_labels;
    sv_itr->getNeighborLabels (neighbor_labels);
    for (const unsigned int &neighbor_label : neighbor_labels)
      label_adjacency.insert (std::pair<std::uint32_t,std::uint32_t> (label, neighbor_label) );
    //if (neighbor_labels.size () == 0)
    //  std::cout << label<<"(size="<<sv_itr->size () << ") has "<<neighbor_labels.size () << "\n";
  }
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> typename pcl::PointCloud<PointT>::Ptr
pcl::SupervoxelClustering<PointT>::getVoxelCentroidCloud () const
{
  typename PointCloudT::Ptr centroid_copy (new PointCloudT);
  copyPointCloud (*voxel_centroid_cloud_, *centroid_copy);
  return centroid_copy;
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> pcl::PointCloud<pcl::PointXYZL>::Ptr
pcl::SupervoxelClustering<PointT>::getLabeledVoxelCloud () const
{
  pcl::PointCloud<pcl::PointXYZL>::Ptr labeled_voxel_cloud (new pcl::PointCloud<pcl::PointXYZL>);
  for (typename HelperListT::const_iterator sv_itr = supervoxel_helpers_.cbegin (); sv_itr != supervoxel_helpers_.cend (); ++sv_itr)
  {
    typename PointCloudT::Ptr voxels;
    sv_itr->getVoxels (voxels);
    pcl::PointCloud<pcl::PointXYZL> xyzl_copy;
    copyPointCloud (*voxels, xyzl_copy);
    
    auto xyzl_copy_itr = xyzl_copy.begin ();
    for ( ; xyzl_copy_itr != xyzl_copy.end (); ++xyzl_copy_itr) 
      xyzl_copy_itr->label = sv_itr->getLabel ();
    
    *labeled_voxel_cloud += xyzl_copy;
  }
  
  return labeled_voxel_cloud;  
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> pcl::PointCloud<pcl::PointXYZL>::Ptr
pcl::SupervoxelClustering<PointT>::getLabeledCloud () const
{
  pcl::PointCloud<pcl::PointXYZL>::Ptr labeled_cloud (new pcl::PointCloud<pcl::PointXYZL>);
  pcl::copyPointCloud (*input_,*labeled_cloud);
  
  auto i_input = input_->begin ();
  for (auto i_labeled = labeled_cloud->begin (); i_labeled != labeled_cloud->end (); ++i_labeled,++i_input)
  {
    if ( !pcl::isFinite<PointT> (*i_input))
      i_labeled->label = 0;
    else
    {     
      i_labeled->label = 0;
      LeafContainerT *leaf = adjacency_octree_->getLeafContainerAtPoint (*i_input);
      VoxelData& voxel_data = leaf->getData ();
      if (voxel_data.owner_)
        i_labeled->label = voxel_data.owner_->getLabel ();
        
    }
      
  }
    
  return (labeled_cloud);
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> pcl::PointCloud<pcl::PointNormal>::Ptr
pcl::SupervoxelClustering<PointT>::makeSupervoxelNormalCloud (std::map<std::uint32_t,typename Supervoxel<PointT>::Ptr > &supervoxel_clusters)
{
  pcl::PointCloud<pcl::PointNormal>::Ptr normal_cloud (new pcl::PointCloud<pcl::PointNormal>);
  normal_cloud->resize (supervoxel_clusters.size ());
  auto normal_cloud_itr = normal_cloud->begin ();
  for (auto sv_itr = supervoxel_clusters.cbegin (), sv_itr_end = supervoxel_clusters.cend ();
       sv_itr != sv_itr_end; ++sv_itr, ++normal_cloud_itr)
  {
    (sv_itr->second)->getCentroidPointNormal (*normal_cloud_itr);
  }
  return normal_cloud;
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> float
pcl::SupervoxelClustering<PointT>::getVoxelResolution () const
{
  return (resolution_);
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::setVoxelResolution (float resolution)
{
  resolution_ = resolution;
  
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> float
pcl::SupervoxelClustering<PointT>::getSeedResolution () const
{
  return (seed_resolution_);
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::setSeedResolution (float seed_resolution)
{
  seed_resolution_ = seed_resolution;
}
 
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::setColorImportance (float val)
{
  color_importance_ = val;
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::setSpatialImportance (float val)
{
  spatial_importance_ = val;
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::setNormalImportance (float val)
{
  normal_importance_ = val;
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::setUseSingleCameraTransform (bool val)
{
  use_default_transform_behaviour_ = false;
  use_single_camera_transform_ = val;
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> int
pcl::SupervoxelClustering<PointT>::getMaxLabel () const
{
  int max_label = 0;
  for (typename HelperListT::const_iterator sv_itr = supervoxel_helpers_.cbegin (); sv_itr != supervoxel_helpers_.cend (); ++sv_itr)
  {
    int temp = sv_itr->getLabel ();
    if (temp > max_label)
      max_label = temp;
  }
  return max_label;
}
 
namespace pcl
{ 
  namespace octree
  {
    //Explicit overloads for RGB types
    template<>
    void
    pcl::octree::OctreePointCloudAdjacencyContainer<pcl::PointXYZRGB,pcl::SupervoxelClustering<pcl::PointXYZRGB>::VoxelData>::addPoint (const pcl::PointXYZRGB &new_point);
    
    template<>
    void
    pcl::octree::OctreePointCloudAdjacencyContainer<pcl::PointXYZRGBA,pcl::SupervoxelClustering<pcl::PointXYZRGBA>::VoxelData>::addPoint (const pcl::PointXYZRGBA &new_point);
    
    //Explicit overloads for RGB types
    template<> void
    pcl::octree::OctreePointCloudAdjacencyContainer<pcl::PointXYZRGB,pcl::SupervoxelClustering<pcl::PointXYZRGB>::VoxelData>::computeData ();
    
    template<> void
    pcl::octree::OctreePointCloudAdjacencyContainer<pcl::PointXYZRGBA,pcl::SupervoxelClustering<pcl::PointXYZRGBA>::VoxelData>::computeData ();
    
    //Explicit overloads for XYZ types
    template<>
    void
    pcl::octree::OctreePointCloudAdjacencyContainer<pcl::PointXYZ,pcl::SupervoxelClustering<pcl::PointXYZ>::VoxelData>::addPoint (const pcl::PointXYZ &new_point);
    
    template<> void
    pcl::octree::OctreePointCloudAdjacencyContainer<pcl::PointXYZ,pcl::SupervoxelClustering<pcl::PointXYZ>::VoxelData>::computeData ();
  }
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
namespace pcl
{
  
  template<> void
  pcl::SupervoxelClustering<pcl::PointXYZRGB>::VoxelData::getPoint (pcl::PointXYZRGB &point_arg) const;
  
  template<> void
  pcl::SupervoxelClustering<pcl::PointXYZRGBA>::VoxelData::getPoint (pcl::PointXYZRGBA &point_arg ) const;
  
  template<typename PointT> void
  pcl::SupervoxelClustering<PointT>::VoxelData::getPoint (PointT &point_arg ) const
  {
    //XYZ is required or this doesn't make much sense...
    point_arg.x = xyz_[0];
    point_arg.y = xyz_[1];
    point_arg.z = xyz_[2];
  }
  
  //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
  template <typename PointT> void
  pcl::SupervoxelClustering<PointT>::VoxelData::getNormal (Normal &normal_arg) const
  {
    normal_arg.normal_x = normal_[0];
    normal_arg.normal_y = normal_[1];
    normal_arg.normal_z = normal_[2];
    normal_arg.curvature = curvature_;
  }
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::SupervoxelHelper::addLeaf (LeafContainerT* leaf_arg)
{
  leaves_.insert (leaf_arg);
  VoxelData& voxel_data = leaf_arg->getData ();
  voxel_data.owner_ = this;
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::SupervoxelHelper::removeLeaf (LeafContainerT* leaf_arg)
{
  leaves_.erase (leaf_arg);
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::SupervoxelHelper::removeAllLeaves ()
{
  for (auto leaf_itr = leaves_.cbegin (); leaf_itr != leaves_.cend (); ++leaf_itr)
  {
    VoxelData& voxel = ((*leaf_itr)->getData ());
    voxel.owner_ = nullptr;
    voxel.distance_ = std::numeric_limits<float>::max ();
  }
  leaves_.clear ();
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::SupervoxelHelper::expand ()
{
  //std::cout << "Expanding sv "<<label_<<", owns "<<leaves_.size ()<<" voxels\n";
  //Buffer of new neighbors - initial size is just a guess of most possible
  std::vector<LeafContainerT*> new_owned;
  new_owned.reserve (leaves_.size () * 9);
  //For each leaf belonging to this supervoxel
  for (auto leaf_itr = leaves_.cbegin (); leaf_itr != leaves_.cend (); ++leaf_itr)
  {
    //for each neighbor of the leaf
    for (auto neighb_itr=(*leaf_itr)->cbegin (); neighb_itr!=(*leaf_itr)->cend (); ++neighb_itr)
    {
      //Get a reference to the data contained in the leaf
      VoxelData& neighbor_voxel = ((*neighb_itr)->getData ());
      //TODO this is a shortcut, really we should always recompute distance
      if(neighbor_voxel.owner_ == this)
        continue;
      //Compute distance to the neighbor
      float dist = parent_->voxelDataDistance (centroid_, neighbor_voxel);
      //If distance is less than previous, we remove it from its owner's list
      //and change the owner to this and distance (we *steal* it!)
      if (dist < neighbor_voxel.distance_)  
      {
        neighbor_voxel.distance_ = dist;
        if (neighbor_voxel.owner_ != this)
        {
          if (neighbor_voxel.owner_)
            (neighbor_voxel.owner_)->removeLeaf(*neighb_itr);
          neighbor_voxel.owner_ = this;
          new_owned.push_back (*neighb_itr);
        }
      }
    }
  }
  //Push all new owned onto the owned leaf set
  for (auto new_owned_itr = new_owned.cbegin (); new_owned_itr != new_owned.cend (); ++new_owned_itr)
  {
    leaves_.insert (*new_owned_itr);
  }  
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::SupervoxelHelper::refineNormals ()
{
  //For each leaf belonging to this supervoxel, get its neighbors, build an index vector, compute normal
  for (auto leaf_itr = leaves_.cbegin (); leaf_itr != leaves_.cend (); ++leaf_itr)
  {
    VoxelData& voxel_data = (*leaf_itr)->getData ();
    Indices indices;
    indices.reserve (81); 
    //Push this point
    indices.push_back (voxel_data.idx_);
    for (auto neighb_itr=(*leaf_itr)->cbegin (); neighb_itr!=(*leaf_itr)->cend (); ++neighb_itr)
    {
      //Get a reference to the data contained in the leaf
      VoxelData& neighbor_voxel_data = ((*neighb_itr)->getData ());
      //If the neighbor is in this supervoxel, use it
      if (neighbor_voxel_data.owner_ == this)
      {
        indices.push_back (neighbor_voxel_data.idx_);
        //Also check its neighbors
        for (auto neighb_neighb_itr=(*neighb_itr)->cbegin (); neighb_neighb_itr!=(*neighb_itr)->cend (); ++neighb_neighb_itr)
        {
          VoxelData& neighb_neighb_voxel_data = (*neighb_neighb_itr)->getData ();
          if (neighb_neighb_voxel_data.owner_ == this)
            indices.push_back (neighb_neighb_voxel_data.idx_);
        }
        
        
      }
    }
    //Compute normal
    pcl::computePointNormal (*parent_->voxel_centroid_cloud_, indices, voxel_data.normal_, voxel_data.curvature_);
    pcl::flipNormalTowardsViewpoint ((*parent_->voxel_centroid_cloud_)[voxel_data.idx_], 0.0f,0.0f,0.0f, voxel_data.normal_);
    voxel_data.normal_[3] = 0.0f;
    voxel_data.normal_.normalize ();
  }
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::SupervoxelHelper::updateCentroid ()
{
  centroid_.normal_ = Eigen::Vector4f::Zero ();
  centroid_.xyz_ = Eigen::Vector3f::Zero ();
  centroid_.rgb_ = Eigen::Vector3f::Zero ();
  for (auto leaf_itr = leaves_.cbegin (); leaf_itr != leaves_.cend (); ++leaf_itr)
  {
    const VoxelData& leaf_data = (*leaf_itr)->getData ();
    centroid_.normal_ += leaf_data.normal_;
    centroid_.xyz_ += leaf_data.xyz_;
    centroid_.rgb_ += leaf_data.rgb_;
  }
  centroid_.normal_.normalize ();
  centroid_.xyz_ /= static_cast<float> (leaves_.size ());
  centroid_.rgb_ /= static_cast<float> (leaves_.size ());  
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::SupervoxelHelper::getVoxels (typename pcl::PointCloud<PointT>::Ptr &voxels) const
{
  voxels.reset (new pcl::PointCloud<PointT>);
  voxels->clear ();
  voxels->resize (leaves_.size ());
  auto voxel_itr = voxels->begin ();
  for (auto leaf_itr = leaves_.cbegin (); leaf_itr != leaves_.cend (); ++leaf_itr, ++voxel_itr)
  {
    const VoxelData& leaf_data = (*leaf_itr)->getData ();
    leaf_data.getPoint (*voxel_itr);
  }
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::SupervoxelHelper::getNormals (typename pcl::PointCloud<Normal>::Ptr &normals) const
{
  normals.reset (new pcl::PointCloud<Normal>);
  normals->clear ();
  normals->resize (leaves_.size ());
  auto normal_itr = normals->begin ();
  for (auto leaf_itr = leaves_.cbegin (); leaf_itr != leaves_.cend (); ++leaf_itr, ++normal_itr)
  {
    const VoxelData& leaf_data = (*leaf_itr)->getData ();
    leaf_data.getNormal (*normal_itr);
  }
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename PointT> void
pcl::SupervoxelClustering<PointT>::SupervoxelHelper::getNeighborLabels (std::set<std::uint32_t> &neighbor_labels) const
{
  neighbor_labels.clear ();
  //For each leaf belonging to this supervoxel
  for (auto leaf_itr = leaves_.cbegin (); leaf_itr != leaves_.cend (); ++leaf_itr)
  {
    //for each neighbor of the leaf
    for (auto neighb_itr=(*leaf_itr)->cbegin (); neighb_itr!=(*leaf_itr)->cend (); ++neighb_itr)
    {
      //Get a reference to the data contained in the leaf
      VoxelData& neighbor_voxel = ((*neighb_itr)->getData ());
      //If it has an owner, and it's not us - get it's owner's label insert into set
      if (neighbor_voxel.owner_ != this && neighbor_voxel.owner_)
      {
        neighbor_labels.insert (neighbor_voxel.owner_->getLabel ());
      }
    }
  }
}
 
 
#endif    // PCL_SUPERVOXEL_CLUSTERING_HPP_