38 #ifndef PCL_SEGMENTATION_IMPL_CPC_SEGMENTATION_HPP_
39 #define PCL_SEGMENTATION_IMPL_CPC_SEGMENTATION_HPP_
41 #include <pcl/segmentation/cpc_segmentation.h>
43 template <
typename Po
intT>
46 min_segment_size_for_cutting_ (400),
47 min_cut_score_ (0.16),
48 use_local_constrains_ (true),
49 use_directed_weights_ (true),
54 template <
typename Po
intT>
59 template <
typename Po
intT>
void
66 calculateConvexConnections (sv_adjacency_list_);
69 applyKconvexity (k_factor_);
74 grouping_data_valid_ =
true;
76 applyCuttingPlane (max_cuts_);
79 mergeSmallSegments ();
82 PCL_WARN (
"[pcl::CPCSegmentation::segment] WARNING: Call function setInputSupervoxels first. Nothing has been done. \n");
85 template <
typename Po
intT>
void
88 using SegLabel2ClusterMap = std::map<std::uint32_t, pcl::PointCloud<WeightSACPointType>::Ptr>;
92 if (depth_levels_left <= 0)
96 SegLabel2ClusterMap seg_to_edge_points_map;
97 std::map<std::uint32_t, std::vector<EdgeID> > seg_to_edgeIDs_map;
98 EdgeIterator edge_itr, edge_itr_end, next_edge;
99 boost::tie (edge_itr, edge_itr_end) = boost::edges (sv_adjacency_list_);
100 for (next_edge = edge_itr; edge_itr != edge_itr_end; edge_itr = next_edge)
103 std::uint32_t source_sv_label = sv_adjacency_list_[boost::source (*edge_itr, sv_adjacency_list_)];
104 std::uint32_t target_sv_label = sv_adjacency_list_[boost::target (*edge_itr, sv_adjacency_list_)];
106 std::uint32_t source_segment_label = sv_label_to_seg_label_map_[source_sv_label];
107 std::uint32_t target_segment_label = sv_label_to_seg_label_map_[target_sv_label];
110 if (source_segment_label != target_segment_label)
114 if (sv_adjacency_list_[*edge_itr].used_for_cutting)
117 const pcl::PointXYZRGBA source_centroid = sv_label_to_supervoxel_map_[source_sv_label]->centroid_;
118 const pcl::PointXYZRGBA target_centroid = sv_label_to_supervoxel_map_[target_sv_label]->centroid_;
123 WeightSACPointType edge_centroid;
124 edge_centroid.getVector3fMap () = (source_centroid.getVector3fMap () + target_centroid.getVector3fMap ()) / 2;
127 edge_centroid.getNormalVector3fMap () = (target_centroid.getVector3fMap () - source_centroid.getVector3fMap ()).normalized ();
130 edge_centroid.intensity = sv_adjacency_list_[*edge_itr].is_convex ? -sv_adjacency_list_[*edge_itr].normal_difference : sv_adjacency_list_[*edge_itr].normal_difference;
131 if (seg_to_edge_points_map.find (source_segment_label) == seg_to_edge_points_map.end ())
135 seg_to_edge_points_map[source_segment_label]->push_back (edge_centroid);
136 seg_to_edgeIDs_map[source_segment_label].push_back (*edge_itr);
138 bool cut_found =
false;
140 for (
const auto &seg_to_edge_points : seg_to_edge_points_map)
143 if (seg_to_edge_points.second->size () < min_segment_size_for_cutting_)
148 std::vector<double> weights;
149 weights.resize (seg_to_edge_points.second->size ());
150 for (std::size_t cp = 0;
cp < seg_to_edge_points.second->size (); ++
cp)
152 float& cur_weight = seg_to_edge_points.second->points[
cp].intensity;
153 cur_weight = cur_weight < concavity_tolerance_threshold_ ? 0 : 1;
154 weights[
cp] = cur_weight;
160 WeightedRandomSampleConsensus weight_sac (model_p, seed_resolution_,
true);
162 weight_sac.setWeights (weights, use_directed_weights_);
163 weight_sac.setMaxIterations (ransac_itrs_);
166 if (!weight_sac.computeModel ())
171 Eigen::VectorXf model_coefficients;
172 weight_sac.getModelCoefficients (model_coefficients);
174 model_coefficients[3] += std::numeric_limits<float>::epsilon ();
176 weight_sac.getInliers (*support_indices);
181 if (use_local_constrains_)
183 Eigen::Vector3f plane_normal (model_coefficients[0], model_coefficients[1], model_coefficients[2]);
187 std::vector<pcl::PointIndices> cluster_indices;
190 tree->setInputCloud (edge_cloud_cluster);
196 euclidean_clusterer.
setIndices (support_indices);
197 euclidean_clusterer.
extract (cluster_indices);
200 for (
const auto &cluster_index : cluster_indices)
203 int cluster_concave_pts = 0;
204 float cluster_score = 0;
206 for (
const int ¤t_index : cluster_index.indices)
208 double index_score = weights[current_index];
209 if (use_directed_weights_)
210 index_score *= 1.414 * (std::abs (plane_normal.dot (edge_cloud_cluster->
at (current_index).getNormalVector3fMap ())));
211 cluster_score += index_score;
212 if (weights[current_index] > 0)
213 ++cluster_concave_pts;
216 cluster_score /= cluster_index.indices.size ();
218 if (cluster_score >= min_cut_score_)
220 cut_support_indices.insert (cut_support_indices.end (), cluster_index.indices.begin (), cluster_index.indices.end ());
223 if (cut_support_indices.empty ())
231 double current_score = weight_sac.getBestScore ();
232 cut_support_indices = *support_indices;
234 if (current_score < min_cut_score_)
241 int number_connections_cut = 0;
242 for (
const int &point_index : cut_support_indices)
244 if (use_clean_cutting_)
247 std::uint32_t source_sv_label = sv_adjacency_list_[boost::source (seg_to_edgeIDs_map[seg_to_edge_points.first][point_index], sv_adjacency_list_)];
248 std::uint32_t target_sv_label = sv_adjacency_list_[boost::target (seg_to_edgeIDs_map[seg_to_edge_points.first][point_index], sv_adjacency_list_)];
250 const pcl::PointXYZRGBA source_centroid = sv_label_to_supervoxel_map_[source_sv_label]->centroid_;
251 const pcl::PointXYZRGBA target_centroid = sv_label_to_supervoxel_map_[target_sv_label]->centroid_;
258 sv_adjacency_list_[seg_to_edgeIDs_map[seg_to_edge_points.first][point_index]].used_for_cutting =
true;
259 if (sv_adjacency_list_[seg_to_edgeIDs_map[seg_to_edge_points.first][point_index]].is_valid)
261 ++number_connections_cut;
262 sv_adjacency_list_[seg_to_edgeIDs_map[seg_to_edge_points.first][point_index]].is_valid =
false;
266 if (number_connections_cut > 0)
275 applyCuttingPlane (depth_levels_left);
284 template <
typename Po
intT>
bool
288 if (threshold_ == std::numeric_limits<double>::max ())
290 PCL_ERROR (
"[pcl::CPCSegmentation<PointT>::WeightedRandomSampleConsensus::computeModel] No threshold set!\n");
295 best_score_ = -std::numeric_limits<double>::max ();
297 std::vector<int> selection;
298 Eigen::VectorXf model_coefficients;
300 unsigned skipped_count = 0;
302 const unsigned max_skip = max_iterations_ * 10;
305 while (iterations_ < max_iterations_ && skipped_count < max_skip)
308 sac_model_->setIndices (model_pt_indices_);
309 sac_model_->getSamples (iterations_, selection);
311 if (selection.empty ())
313 PCL_ERROR (
"[pcl::CPCSegmentation<PointT>::WeightedRandomSampleConsensus::computeModel] No samples could be selected!\n");
318 if (!sac_model_->computeModelCoefficients (selection, model_coefficients))
325 sac_model_->setIndices (full_cloud_pt_indices_);
328 sac_model_->selectWithinDistance (model_coefficients, threshold_, *current_inliers);
329 double current_score = 0;
330 Eigen::Vector3f plane_normal (model_coefficients[0], model_coefficients[1], model_coefficients[2]);
331 for (
const int ¤t_index : *current_inliers)
333 double index_score = weights_[current_index];
334 if (use_directed_weights_)
336 index_score *= 1.414 * (std::abs (plane_normal.dot (point_cloud_ptr_->at (current_index).getNormalVector3fMap ())));
338 current_score += index_score;
341 current_score /= current_inliers->size ();
344 if (current_score > best_score_)
346 best_score_ = current_score;
349 model_coefficients_ = model_coefficients;
353 PCL_DEBUG (
"[pcl::CPCSegmentation<PointT>::WeightedRandomSampleConsensus::computeModel] Trial %d (max %d): score is %f (best is: %f so far).\n", iterations_, max_iterations_, current_score, best_score_);
354 if (iterations_ > max_iterations_)
356 PCL_DEBUG (
"[pcl::CPCSegmentation<PointT>::WeightedRandomSampleConsensus::computeModel] RANSAC reached the maximum number of trials.\n");
361 PCL_DEBUG (
"[pcl::CPCSegmentation<PointT>::WeightedRandomSampleConsensus::computeModel] Model: %lu size, %f score.\n", model_.size (), best_score_);
370 sac_model_->selectWithinDistance (model_coefficients_, threshold_, inliers_);
374 #endif // PCL_SEGMENTATION_IMPL_CPC_SEGMENTATION_HPP_