1 // Ceres Solver - A fast non-linear least squares minimizer
2 // Copyright 2010, 2011, 2012 Google Inc. All rights reserved.
3 // http://code.google.com/p/ceres-solver/
4 //
5 // Redistribution and use in source and binary forms, with or without
6 // modification, are permitted provided that the following conditions are met:
7 //
8 // * Redistributions of source code must retain the above copyright notice,
9 //   this list of conditions and the following disclaimer.
10 // * Redistributions in binary form must reproduce the above copyright notice,
11 //   this list of conditions and the following disclaimer in the documentation
12 //   and/or other materials provided with the distribution.
13 // * Neither the name of Google Inc. nor the names of its contributors may be
14 //   used to endorse or promote products derived from this software without
15 //   specific prior written permission.
16 //
17 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
18 // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 // ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
21 // LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
22 // CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
23 // SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
24 // INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
25 // CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
26 // ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27 // POSSIBILITY OF SUCH DAMAGE.
28 //
29 // Author: sameeragarwal@google.com (Sameer Agarwal)
30 
31 // This include must come before any #ifndef check on Ceres compile options.
32 #include "ceres/internal/port.h"
33 
34 #ifndef CERES_NO_SUITESPARSE
35 
36 #include "ceres/visibility_based_preconditioner.h"
37 
38 #include <algorithm>
39 #include <functional>
40 #include <iterator>
41 #include <set>
42 #include <utility>
43 #include <vector>
44 #include "Eigen/Dense"
45 #include "ceres/block_random_access_sparse_matrix.h"
46 #include "ceres/block_sparse_matrix.h"
47 #include "ceres/canonical_views_clustering.h"
48 #include "ceres/collections_port.h"
49 #include "ceres/graph.h"
50 #include "ceres/graph_algorithms.h"
51 #include "ceres/internal/scoped_ptr.h"
52 #include "ceres/linear_solver.h"
53 #include "ceres/schur_eliminator.h"
54 #include "ceres/single_linkage_clustering.h"
55 #include "ceres/visibility.h"
56 #include "glog/logging.h"
57 
58 namespace ceres {
59 namespace internal {
60 
61 // TODO(sameeragarwal): Currently these are magic weights for the
62 // preconditioner construction. Move these higher up into the Options
63 // struct and provide some guidelines for choosing them.
64 //
65 // This will require some more work on the clustering algorithm and
66 // possibly some more refactoring of the code.
67 static const double kCanonicalViewsSizePenaltyWeight = 3.0;
68 static const double kCanonicalViewsSimilarityPenaltyWeight = 0.0;
69 static const double kSingleLinkageMinSimilarity = 0.9;
70 
VisibilityBasedPreconditioner(const CompressedRowBlockStructure & bs,const Preconditioner::Options & options)71 VisibilityBasedPreconditioner::VisibilityBasedPreconditioner(
72     const CompressedRowBlockStructure& bs,
73     const Preconditioner::Options& options)
74     : options_(options),
75       num_blocks_(0),
76       num_clusters_(0),
77       factor_(NULL) {
78   CHECK_GT(options_.elimination_groups.size(), 1);
79   CHECK_GT(options_.elimination_groups[0], 0);
80   CHECK(options_.type == CLUSTER_JACOBI ||
81         options_.type == CLUSTER_TRIDIAGONAL)
82       << "Unknown preconditioner type: " << options_.type;
83   num_blocks_ = bs.cols.size() - options_.elimination_groups[0];
84   CHECK_GT(num_blocks_, 0)
85       << "Jacobian should have atleast 1 f_block for "
86       << "visibility based preconditioning.";
87 
88   // Vector of camera block sizes
89   block_size_.resize(num_blocks_);
90   for (int i = 0; i < num_blocks_; ++i) {
91     block_size_[i] = bs.cols[i + options_.elimination_groups[0]].size;
92   }
93 
94   const time_t start_time = time(NULL);
95   switch (options_.type) {
96     case CLUSTER_JACOBI:
97       ComputeClusterJacobiSparsity(bs);
98       break;
99     case CLUSTER_TRIDIAGONAL:
100       ComputeClusterTridiagonalSparsity(bs);
101       break;
102     default:
103       LOG(FATAL) << "Unknown preconditioner type";
104   }
105   const time_t structure_time = time(NULL);
106   InitStorage(bs);
107   const time_t storage_time = time(NULL);
108   InitEliminator(bs);
109   const time_t eliminator_time = time(NULL);
110 
111   // Allocate temporary storage for a vector used during
112   // RightMultiply.
113   tmp_rhs_ = CHECK_NOTNULL(ss_.CreateDenseVector(NULL,
114                                                  m_->num_rows(),
115                                                  m_->num_rows()));
116   const time_t init_time = time(NULL);
117   VLOG(2) << "init time: "
118           << init_time - start_time
119           << " structure time: " << structure_time - start_time
120           << " storage time:" << storage_time - structure_time
121           << " eliminator time: " << eliminator_time - storage_time;
122 }
123 
~VisibilityBasedPreconditioner()124 VisibilityBasedPreconditioner::~VisibilityBasedPreconditioner() {
125   if (factor_ != NULL) {
126     ss_.Free(factor_);
127     factor_ = NULL;
128   }
129   if (tmp_rhs_ != NULL) {
130     ss_.Free(tmp_rhs_);
131     tmp_rhs_ = NULL;
132   }
133 }
134 
135 // Determine the sparsity structure of the CLUSTER_JACOBI
136 // preconditioner. It clusters cameras using their scene
137 // visibility. The clusters form the diagonal blocks of the
138 // preconditioner matrix.
ComputeClusterJacobiSparsity(const CompressedRowBlockStructure & bs)139 void VisibilityBasedPreconditioner::ComputeClusterJacobiSparsity(
140     const CompressedRowBlockStructure& bs) {
141   vector<set<int> > visibility;
142   ComputeVisibility(bs, options_.elimination_groups[0], &visibility);
143   CHECK_EQ(num_blocks_, visibility.size());
144   ClusterCameras(visibility);
145   cluster_pairs_.clear();
146   for (int i = 0; i < num_clusters_; ++i) {
147     cluster_pairs_.insert(make_pair(i, i));
148   }
149 }
150 
151 // Determine the sparsity structure of the CLUSTER_TRIDIAGONAL
152 // preconditioner. It clusters cameras using using the scene
153 // visibility and then finds the strongly interacting pairs of
154 // clusters by constructing another graph with the clusters as
155 // vertices and approximating it with a degree-2 maximum spanning
156 // forest. The set of edges in this forest are the cluster pairs.
ComputeClusterTridiagonalSparsity(const CompressedRowBlockStructure & bs)157 void VisibilityBasedPreconditioner::ComputeClusterTridiagonalSparsity(
158     const CompressedRowBlockStructure& bs) {
159   vector<set<int> > visibility;
160   ComputeVisibility(bs, options_.elimination_groups[0], &visibility);
161   CHECK_EQ(num_blocks_, visibility.size());
162   ClusterCameras(visibility);
163 
164   // Construct a weighted graph on the set of clusters, where the
165   // edges are the number of 3D points/e_blocks visible in both the
166   // clusters at the ends of the edge. Return an approximate degree-2
167   // maximum spanning forest of this graph.
168   vector<set<int> > cluster_visibility;
169   ComputeClusterVisibility(visibility, &cluster_visibility);
170   scoped_ptr<Graph<int> > cluster_graph(
171       CHECK_NOTNULL(CreateClusterGraph(cluster_visibility)));
172   scoped_ptr<Graph<int> > forest(
173       CHECK_NOTNULL(Degree2MaximumSpanningForest(*cluster_graph)));
174   ForestToClusterPairs(*forest, &cluster_pairs_);
175 }
176 
177 // Allocate storage for the preconditioner matrix.
InitStorage(const CompressedRowBlockStructure & bs)178 void VisibilityBasedPreconditioner::InitStorage(
179     const CompressedRowBlockStructure& bs) {
180   ComputeBlockPairsInPreconditioner(bs);
181   m_.reset(new BlockRandomAccessSparseMatrix(block_size_, block_pairs_));
182 }
183 
184 // Call the canonical views algorithm and cluster the cameras based on
185 // their visibility sets. The visibility set of a camera is the set of
186 // e_blocks/3D points in the scene that are seen by it.
187 //
188 // The cluster_membership_ vector is updated to indicate cluster
189 // memberships for each camera block.
ClusterCameras(const vector<set<int>> & visibility)190 void VisibilityBasedPreconditioner::ClusterCameras(
191     const vector<set<int> >& visibility) {
192   scoped_ptr<Graph<int> > schur_complement_graph(
193       CHECK_NOTNULL(CreateSchurComplementGraph(visibility)));
194 
195   HashMap<int, int> membership;
196 
197   if (options_.visibility_clustering_type == CANONICAL_VIEWS) {
198     vector<int> centers;
199     CanonicalViewsClusteringOptions clustering_options;
200     clustering_options.size_penalty_weight =
201         kCanonicalViewsSizePenaltyWeight;
202     clustering_options.similarity_penalty_weight =
203         kCanonicalViewsSimilarityPenaltyWeight;
204     ComputeCanonicalViewsClustering(clustering_options,
205                                     *schur_complement_graph,
206                                     &centers,
207                                     &membership);
208     num_clusters_ = centers.size();
209   } else if (options_.visibility_clustering_type == SINGLE_LINKAGE) {
210     SingleLinkageClusteringOptions clustering_options;
211     clustering_options.min_similarity =
212         kSingleLinkageMinSimilarity;
213     num_clusters_ = ComputeSingleLinkageClustering(clustering_options,
214                                                    *schur_complement_graph,
215                                                    &membership);
216   } else {
217     LOG(FATAL) << "Unknown visibility clustering algorithm.";
218   }
219 
220   CHECK_GT(num_clusters_, 0);
221   VLOG(2) << "num_clusters: " << num_clusters_;
222   FlattenMembershipMap(membership, &cluster_membership_);
223 }
224 
225 // Compute the block sparsity structure of the Schur complement
226 // matrix. For each pair of cameras contributing a non-zero cell to
227 // the schur complement, determine if that cell is present in the
228 // preconditioner or not.
229 //
230 // A pair of cameras contribute a cell to the preconditioner if they
231 // are part of the same cluster or if the the two clusters that they
232 // belong have an edge connecting them in the degree-2 maximum
233 // spanning forest.
234 //
235 // For example, a camera pair (i,j) where i belonges to cluster1 and
236 // j belongs to cluster2 (assume that cluster1 < cluster2).
237 //
238 // The cell corresponding to (i,j) is present in the preconditioner
239 // if cluster1 == cluster2 or the pair (cluster1, cluster2) were
240 // connected by an edge in the degree-2 maximum spanning forest.
241 //
242 // Since we have already expanded the forest into a set of camera
243 // pairs/edges, including self edges, the check can be reduced to
244 // checking membership of (cluster1, cluster2) in cluster_pairs_.
ComputeBlockPairsInPreconditioner(const CompressedRowBlockStructure & bs)245 void VisibilityBasedPreconditioner::ComputeBlockPairsInPreconditioner(
246     const CompressedRowBlockStructure& bs) {
247   block_pairs_.clear();
248   for (int i = 0; i < num_blocks_; ++i) {
249     block_pairs_.insert(make_pair(i, i));
250   }
251 
252   int r = 0;
253   const int num_row_blocks = bs.rows.size();
254   const int num_eliminate_blocks = options_.elimination_groups[0];
255 
256   // Iterate over each row of the matrix. The block structure of the
257   // matrix is assumed to be sorted in order of the e_blocks/point
258   // blocks. Thus all row blocks containing an e_block/point occur
259   // contiguously. Further, if present, an e_block is always the first
260   // parameter block in each row block.  These structural assumptions
261   // are common to all Schur complement based solvers in Ceres.
262   //
263   // For each e_block/point block we identify the set of cameras
264   // seeing it. The cross product of this set with itself is the set
265   // of non-zero cells contibuted by this e_block.
266   //
267   // The time complexity of this is O(nm^2) where, n is the number of
268   // 3d points and m is the maximum number of cameras seeing any
269   // point, which for most scenes is a fairly small number.
270   while (r < num_row_blocks) {
271     int e_block_id = bs.rows[r].cells.front().block_id;
272     if (e_block_id >= num_eliminate_blocks) {
273       // Skip the rows whose first block is an f_block.
274       break;
275     }
276 
277     set<int> f_blocks;
278     for (; r < num_row_blocks; ++r) {
279       const CompressedRow& row = bs.rows[r];
280       if (row.cells.front().block_id != e_block_id) {
281         break;
282       }
283 
284       // Iterate over the blocks in the row, ignoring the first block
285       // since it is the one to be eliminated and adding the rest to
286       // the list of f_blocks associated with this e_block.
287       for (int c = 1; c < row.cells.size(); ++c) {
288         const Cell& cell = row.cells[c];
289         const int f_block_id = cell.block_id - num_eliminate_blocks;
290         CHECK_GE(f_block_id, 0);
291         f_blocks.insert(f_block_id);
292       }
293     }
294 
295     for (set<int>::const_iterator block1 = f_blocks.begin();
296          block1 != f_blocks.end();
297          ++block1) {
298       set<int>::const_iterator block2 = block1;
299       ++block2;
300       for (; block2 != f_blocks.end(); ++block2) {
301         if (IsBlockPairInPreconditioner(*block1, *block2)) {
302           block_pairs_.insert(make_pair(*block1, *block2));
303         }
304       }
305     }
306   }
307 
308   // The remaining rows which do not contain any e_blocks.
309   for (; r < num_row_blocks; ++r) {
310     const CompressedRow& row = bs.rows[r];
311     CHECK_GE(row.cells.front().block_id, num_eliminate_blocks);
312     for (int i = 0; i < row.cells.size(); ++i) {
313       const int block1 = row.cells[i].block_id - num_eliminate_blocks;
314       for (int j = 0; j < row.cells.size(); ++j) {
315         const int block2 = row.cells[j].block_id - num_eliminate_blocks;
316         if (block1 <= block2) {
317           if (IsBlockPairInPreconditioner(block1, block2)) {
318             block_pairs_.insert(make_pair(block1, block2));
319           }
320         }
321       }
322     }
323   }
324 
325   VLOG(1) << "Block pair stats: " << block_pairs_.size();
326 }
327 
328 // Initialize the SchurEliminator.
InitEliminator(const CompressedRowBlockStructure & bs)329 void VisibilityBasedPreconditioner::InitEliminator(
330     const CompressedRowBlockStructure& bs) {
331   LinearSolver::Options eliminator_options;
332   eliminator_options.elimination_groups = options_.elimination_groups;
333   eliminator_options.num_threads = options_.num_threads;
334   eliminator_options.e_block_size = options_.e_block_size;
335   eliminator_options.f_block_size = options_.f_block_size;
336   eliminator_options.row_block_size = options_.row_block_size;
337   eliminator_.reset(SchurEliminatorBase::Create(eliminator_options));
338   eliminator_->Init(eliminator_options.elimination_groups[0], &bs);
339 }
340 
341 // Update the values of the preconditioner matrix and factorize it.
UpdateImpl(const BlockSparseMatrix & A,const double * D)342 bool VisibilityBasedPreconditioner::UpdateImpl(const BlockSparseMatrix& A,
343                                                const double* D) {
344   const time_t start_time = time(NULL);
345   const int num_rows = m_->num_rows();
346   CHECK_GT(num_rows, 0);
347 
348   // We need a dummy rhs vector and a dummy b vector since the Schur
349   // eliminator combines the computation of the reduced camera matrix
350   // with the computation of the right hand side of that linear
351   // system.
352   //
353   // TODO(sameeragarwal): Perhaps its worth refactoring the
354   // SchurEliminator::Eliminate function to allow NULL for the rhs. As
355   // of now it does not seem to be worth the effort.
356   Vector rhs = Vector::Zero(m_->num_rows());
357   Vector b = Vector::Zero(A.num_rows());
358 
359   // Compute a subset of the entries of the Schur complement.
360   eliminator_->Eliminate(&A, b.data(), D, m_.get(), rhs.data());
361 
362   // Try factorizing the matrix. For CLUSTER_JACOBI, this should
363   // always succeed modulo some numerical/conditioning problems. For
364   // CLUSTER_TRIDIAGONAL, in general the preconditioner matrix as
365   // constructed is not positive definite. However, we will go ahead
366   // and try factorizing it. If it works, great, otherwise we scale
367   // all the cells in the preconditioner corresponding to the edges in
368   // the degree-2 forest and that guarantees positive
369   // definiteness. The proof of this fact can be found in Lemma 1 in
370   // "Visibility Based Preconditioning for Bundle Adjustment".
371   //
372   // Doing the factorization like this saves us matrix mass when
373   // scaling is not needed, which is quite often in our experience.
374   LinearSolverTerminationType status = Factorize();
375 
376   if (status == LINEAR_SOLVER_FATAL_ERROR) {
377     return false;
378   }
379 
380   // The scaling only affects the tri-diagonal case, since
381   // ScaleOffDiagonalBlocks only pays attenion to the cells that
382   // belong to the edges of the degree-2 forest. In the CLUSTER_JACOBI
383   // case, the preconditioner is guaranteed to be positive
384   // semidefinite.
385   if (status == LINEAR_SOLVER_FAILURE && options_.type == CLUSTER_TRIDIAGONAL) {
386     VLOG(1) << "Unscaled factorization failed. Retrying with off-diagonal "
387             << "scaling";
388     ScaleOffDiagonalCells();
389     status = Factorize();
390   }
391 
392   VLOG(2) << "Compute time: " << time(NULL) - start_time;
393   return (status == LINEAR_SOLVER_SUCCESS);
394 }
395 
396 // Consider the preconditioner matrix as meta-block matrix, whose
397 // blocks correspond to the clusters. Then cluster pairs corresponding
398 // to edges in the degree-2 forest are off diagonal entries of this
399 // matrix. Scaling these off-diagonal entries by 1/2 forces this
400 // matrix to be positive definite.
ScaleOffDiagonalCells()401 void VisibilityBasedPreconditioner::ScaleOffDiagonalCells() {
402   for (set< pair<int, int> >::const_iterator it = block_pairs_.begin();
403        it != block_pairs_.end();
404        ++it) {
405     const int block1 = it->first;
406     const int block2 = it->second;
407     if (!IsBlockPairOffDiagonal(block1, block2)) {
408       continue;
409     }
410 
411     int r, c, row_stride, col_stride;
412     CellInfo* cell_info = m_->GetCell(block1, block2,
413                                       &r, &c,
414                                       &row_stride, &col_stride);
415     CHECK(cell_info != NULL)
416         << "Cell missing for block pair (" << block1 << "," << block2 << ")"
417         << " cluster pair (" << cluster_membership_[block1]
418         << " " << cluster_membership_[block2] << ")";
419 
420     // Ah the magic of tri-diagonal matrices and diagonal
421     // dominance. See Lemma 1 in "Visibility Based Preconditioning
422     // For Bundle Adjustment".
423     MatrixRef m(cell_info->values, row_stride, col_stride);
424     m.block(r, c, block_size_[block1], block_size_[block2]) *= 0.5;
425   }
426 }
427 
428 // Compute the sparse Cholesky factorization of the preconditioner
429 // matrix.
Factorize()430 LinearSolverTerminationType VisibilityBasedPreconditioner::Factorize() {
431   // Extract the TripletSparseMatrix that is used for actually storing
432   // S and convert it into a cholmod_sparse object.
433   cholmod_sparse* lhs = ss_.CreateSparseMatrix(
434       down_cast<BlockRandomAccessSparseMatrix*>(
435           m_.get())->mutable_matrix());
436 
437   // The matrix is symmetric, and the upper triangular part of the
438   // matrix contains the values.
439   lhs->stype = 1;
440 
441   // TODO(sameeragarwal): Refactor to pipe this up and out.
442   string status;
443 
444   // Symbolic factorization is computed if we don't already have one handy.
445   if (factor_ == NULL) {
446     factor_ = ss_.BlockAnalyzeCholesky(lhs, block_size_, block_size_, &status);
447   }
448 
449   const LinearSolverTerminationType termination_type =
450       (factor_ != NULL)
451       ? ss_.Cholesky(lhs, factor_, &status)
452       : LINEAR_SOLVER_FATAL_ERROR;
453 
454   ss_.Free(lhs);
455   return termination_type;
456 }
457 
RightMultiply(const double * x,double * y) const458 void VisibilityBasedPreconditioner::RightMultiply(const double* x,
459                                                   double* y) const {
460   CHECK_NOTNULL(x);
461   CHECK_NOTNULL(y);
462   SuiteSparse* ss = const_cast<SuiteSparse*>(&ss_);
463 
464   const int num_rows = m_->num_rows();
465   memcpy(CHECK_NOTNULL(tmp_rhs_)->x, x, m_->num_rows() * sizeof(*x));
466   // TODO(sameeragarwal): Better error handling.
467   string status;
468   cholmod_dense* solution =
469       CHECK_NOTNULL(ss->Solve(factor_, tmp_rhs_, &status));
470   memcpy(y, solution->x, sizeof(*y) * num_rows);
471   ss->Free(solution);
472 }
473 
num_rows() const474 int VisibilityBasedPreconditioner::num_rows() const {
475   return m_->num_rows();
476 }
477 
478 // Classify camera/f_block pairs as in and out of the preconditioner,
479 // based on whether the cluster pair that they belong to is in the
480 // preconditioner or not.
IsBlockPairInPreconditioner(const int block1,const int block2) const481 bool VisibilityBasedPreconditioner::IsBlockPairInPreconditioner(
482     const int block1,
483     const int block2) const {
484   int cluster1 = cluster_membership_[block1];
485   int cluster2 = cluster_membership_[block2];
486   if (cluster1 > cluster2) {
487     std::swap(cluster1, cluster2);
488   }
489   return (cluster_pairs_.count(make_pair(cluster1, cluster2)) > 0);
490 }
491 
IsBlockPairOffDiagonal(const int block1,const int block2) const492 bool VisibilityBasedPreconditioner::IsBlockPairOffDiagonal(
493     const int block1,
494     const int block2) const {
495   return (cluster_membership_[block1] != cluster_membership_[block2]);
496 }
497 
498 // Convert a graph into a list of edges that includes self edges for
499 // each vertex.
ForestToClusterPairs(const Graph<int> & forest,HashSet<pair<int,int>> * cluster_pairs) const500 void VisibilityBasedPreconditioner::ForestToClusterPairs(
501     const Graph<int>& forest,
502     HashSet<pair<int, int> >* cluster_pairs) const {
503   CHECK_NOTNULL(cluster_pairs)->clear();
504   const HashSet<int>& vertices = forest.vertices();
505   CHECK_EQ(vertices.size(), num_clusters_);
506 
507   // Add all the cluster pairs corresponding to the edges in the
508   // forest.
509   for (HashSet<int>::const_iterator it1 = vertices.begin();
510        it1 != vertices.end();
511        ++it1) {
512     const int cluster1 = *it1;
513     cluster_pairs->insert(make_pair(cluster1, cluster1));
514     const HashSet<int>& neighbors = forest.Neighbors(cluster1);
515     for (HashSet<int>::const_iterator it2 = neighbors.begin();
516          it2 != neighbors.end();
517          ++it2) {
518       const int cluster2 = *it2;
519       if (cluster1 < cluster2) {
520         cluster_pairs->insert(make_pair(cluster1, cluster2));
521       }
522     }
523   }
524 }
525 
526 // The visibilty set of a cluster is the union of the visibilty sets
527 // of all its cameras. In other words, the set of points visible to
528 // any camera in the cluster.
ComputeClusterVisibility(const vector<set<int>> & visibility,vector<set<int>> * cluster_visibility) const529 void VisibilityBasedPreconditioner::ComputeClusterVisibility(
530     const vector<set<int> >& visibility,
531     vector<set<int> >* cluster_visibility) const {
532   CHECK_NOTNULL(cluster_visibility)->resize(0);
533   cluster_visibility->resize(num_clusters_);
534   for (int i = 0; i < num_blocks_; ++i) {
535     const int cluster_id = cluster_membership_[i];
536     (*cluster_visibility)[cluster_id].insert(visibility[i].begin(),
537                                              visibility[i].end());
538   }
539 }
540 
541 // Construct a graph whose vertices are the clusters, and the edge
542 // weights are the number of 3D points visible to cameras in both the
543 // vertices.
CreateClusterGraph(const vector<set<int>> & cluster_visibility) const544 Graph<int>* VisibilityBasedPreconditioner::CreateClusterGraph(
545     const vector<set<int> >& cluster_visibility) const {
546   Graph<int>* cluster_graph = new Graph<int>;
547 
548   for (int i = 0; i < num_clusters_; ++i) {
549     cluster_graph->AddVertex(i);
550   }
551 
552   for (int i = 0; i < num_clusters_; ++i) {
553     const set<int>& cluster_i = cluster_visibility[i];
554     for (int j = i+1; j < num_clusters_; ++j) {
555       vector<int> intersection;
556       const set<int>& cluster_j = cluster_visibility[j];
557       set_intersection(cluster_i.begin(), cluster_i.end(),
558                        cluster_j.begin(), cluster_j.end(),
559                        back_inserter(intersection));
560 
561       if (intersection.size() > 0) {
562         // Clusters interact strongly when they share a large number
563         // of 3D points. The degree-2 maximum spanning forest
564         // alorithm, iterates on the edges in decreasing order of
565         // their weight, which is the number of points shared by the
566         // two cameras that it connects.
567         cluster_graph->AddEdge(i, j, intersection.size());
568       }
569     }
570   }
571   return cluster_graph;
572 }
573 
574 // Canonical views clustering returns a HashMap from vertices to
575 // cluster ids. Convert this into a flat array for quick lookup. It is
576 // possible that some of the vertices may not be associated with any
577 // cluster. In that case, randomly assign them to one of the clusters.
578 //
579 // The cluster ids can be non-contiguous integers. So as we flatten
580 // the membership_map, we also map the cluster ids to a contiguous set
581 // of integers so that the cluster ids are in [0, num_clusters_).
FlattenMembershipMap(const HashMap<int,int> & membership_map,vector<int> * membership_vector) const582 void VisibilityBasedPreconditioner::FlattenMembershipMap(
583     const HashMap<int, int>& membership_map,
584     vector<int>* membership_vector) const {
585   CHECK_NOTNULL(membership_vector)->resize(0);
586   membership_vector->resize(num_blocks_, -1);
587 
588   HashMap<int, int> cluster_id_to_index;
589   // Iterate over the cluster membership map and update the
590   // cluster_membership_ vector assigning arbitrary cluster ids to
591   // the few cameras that have not been clustered.
592   for (HashMap<int, int>::const_iterator it = membership_map.begin();
593        it != membership_map.end();
594        ++it) {
595     const int camera_id = it->first;
596     int cluster_id = it->second;
597 
598     // If the view was not clustered, randomly assign it to one of the
599     // clusters. This preserves the mathematical correctness of the
600     // preconditioner. If there are too many views which are not
601     // clustered, it may lead to some quality degradation though.
602     //
603     // TODO(sameeragarwal): Check if a large number of views have not
604     // been clustered and deal with it?
605     if (cluster_id == -1) {
606       cluster_id = camera_id % num_clusters_;
607     }
608 
609     const int index = FindWithDefault(cluster_id_to_index,
610                                       cluster_id,
611                                       cluster_id_to_index.size());
612 
613     if (index == cluster_id_to_index.size()) {
614       cluster_id_to_index[cluster_id] = index;
615     }
616 
617     CHECK_LT(index, num_clusters_);
618     membership_vector->at(camera_id) = index;
619   }
620 }
621 
622 }  // namespace internal
623 }  // namespace ceres
624 
625 #endif  // CERES_NO_SUITESPARSE
626