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 // Preconditioners for linear systems that arise in Structure from
32 // Motion problems. VisibilityBasedPreconditioner implements:
33 //
34 //  CLUSTER_JACOBI
35 //  CLUSTER_TRIDIAGONAL
36 //
37 // Detailed descriptions of these preconditions beyond what is
38 // documented here can be found in
39 //
40 // Visibility Based Preconditioning for Bundle Adjustment
41 // A. Kushal & S. Agarwal, CVPR 2012.
42 //
43 // http://www.cs.washington.edu/homes/sagarwal/vbp.pdf
44 //
45 // The two preconditioners share enough code that its most efficient
46 // to implement them as part of the same code base.
47 
48 #ifndef CERES_INTERNAL_VISIBILITY_BASED_PRECONDITIONER_H_
49 #define CERES_INTERNAL_VISIBILITY_BASED_PRECONDITIONER_H_
50 
51 #include <set>
52 #include <vector>
53 #include <utility>
54 #include "ceres/collections_port.h"
55 #include "ceres/graph.h"
56 #include "ceres/internal/macros.h"
57 #include "ceres/internal/scoped_ptr.h"
58 #include "ceres/linear_solver.h"
59 #include "ceres/preconditioner.h"
60 #include "ceres/suitesparse.h"
61 
62 namespace ceres {
63 namespace internal {
64 
65 class BlockRandomAccessSparseMatrix;
66 class BlockSparseMatrix;
67 struct CompressedRowBlockStructure;
68 class SchurEliminatorBase;
69 
70 // This class implements visibility based preconditioners for
71 // Structure from Motion/Bundle Adjustment problems. The name
72 // VisibilityBasedPreconditioner comes from the fact that the sparsity
73 // structure of the preconditioner matrix is determined by analyzing
74 // the visibility structure of the scene, i.e. which cameras see which
75 // points.
76 //
77 // The key idea of visibility based preconditioning is to identify
78 // cameras that we expect have strong interactions, and then using the
79 // entries in the Schur complement matrix corresponding to these
80 // camera pairs as an approximation to the full Schur complement.
81 //
82 // CLUSTER_JACOBI identifies these camera pairs by clustering cameras,
83 // and considering all non-zero camera pairs within each cluster. The
84 // clustering in the current implementation is done using the
85 // Canonical Views algorithm of Simon et al. (see
86 // canonical_views_clustering.h). For the purposes of clustering, the
87 // similarity or the degree of interaction between a pair of cameras
88 // is measured by counting the number of points visible in both the
89 // cameras. Thus the name VisibilityBasedPreconditioner. Further, if we
90 // were to permute the parameter blocks such that all the cameras in
91 // the same cluster occur contiguously, the preconditioner matrix will
92 // be a block diagonal matrix with blocks corresponding to the
93 // clusters. Thus in analogy with the Jacobi preconditioner we refer
94 // to this as the CLUSTER_JACOBI preconditioner.
95 //
96 // CLUSTER_TRIDIAGONAL adds more mass to the CLUSTER_JACOBI
97 // preconditioner by considering the interaction between clusters and
98 // identifying strong interactions between cluster pairs. This is done
99 // by constructing a weighted graph on the clusters, with the weight
100 // on the edges connecting two clusters proportional to the number of
101 // 3D points visible to cameras in both the clusters. A degree-2
102 // maximum spanning forest is identified in this graph and the camera
103 // pairs contained in the edges of this forest are added to the
104 // preconditioner. The detailed reasoning for this construction is
105 // explained in the paper mentioned above.
106 //
107 // Degree-2 spanning trees and forests have the property that they
108 // correspond to tri-diagonal matrices. Thus there exist a permutation
109 // of the camera blocks under which the CLUSTER_TRIDIAGONAL
110 // preconditioner matrix is a block tridiagonal matrix, and thus the
111 // name for the preconditioner.
112 //
113 // Thread Safety: This class is NOT thread safe.
114 //
115 // Example usage:
116 //
117 //   LinearSolver::Options options;
118 //   options.preconditioner_type = CLUSTER_JACOBI;
119 //   options.elimination_groups.push_back(num_points);
120 //   options.elimination_groups.push_back(num_cameras);
121 //   VisibilityBasedPreconditioner preconditioner(
122 //      *A.block_structure(), options);
123 //   preconditioner.Update(A, NULL);
124 //   preconditioner.RightMultiply(x, y);
125 //
126 #ifndef CERES_NO_SUITESPARSE
127 class VisibilityBasedPreconditioner : public BlockSparseMatrixPreconditioner {
128  public:
129   // Initialize the symbolic structure of the preconditioner. bs is
130   // the block structure of the linear system to be solved. It is used
131   // to determine the sparsity structure of the preconditioner matrix.
132   //
133   // It has the same structural requirement as other Schur complement
134   // based solvers. Please see schur_eliminator.h for more details.
135   VisibilityBasedPreconditioner(const CompressedRowBlockStructure& bs,
136                                 const Preconditioner::Options& options);
137   virtual ~VisibilityBasedPreconditioner();
138 
139   // Preconditioner interface
140   virtual void RightMultiply(const double* x, double* y) const;
141   virtual int num_rows() const;
142 
143   friend class VisibilityBasedPreconditionerTest;
144 
145  private:
146   virtual bool UpdateImpl(const BlockSparseMatrix& A, const double* D);
147   void ComputeClusterJacobiSparsity(const CompressedRowBlockStructure& bs);
148   void ComputeClusterTridiagonalSparsity(const CompressedRowBlockStructure& bs);
149   void InitStorage(const CompressedRowBlockStructure& bs);
150   void InitEliminator(const CompressedRowBlockStructure& bs);
151   LinearSolverTerminationType Factorize();
152   void ScaleOffDiagonalCells();
153 
154   void ClusterCameras(const vector< set<int> >& visibility);
155   void FlattenMembershipMap(const HashMap<int, int>& membership_map,
156                             vector<int>* membership_vector) const;
157   void ComputeClusterVisibility(const vector<set<int> >& visibility,
158                                 vector<set<int> >* cluster_visibility) const;
159   Graph<int>* CreateClusterGraph(const vector<set<int> >& visibility) const;
160   void ForestToClusterPairs(const Graph<int>& forest,
161                             HashSet<pair<int, int> >* cluster_pairs) const;
162   void ComputeBlockPairsInPreconditioner(const CompressedRowBlockStructure& bs);
163   bool IsBlockPairInPreconditioner(int block1, int block2) const;
164   bool IsBlockPairOffDiagonal(int block1, int block2) const;
165 
166   Preconditioner::Options options_;
167 
168   // Number of parameter blocks in the schur complement.
169   int num_blocks_;
170   int num_clusters_;
171 
172   // Sizes of the blocks in the schur complement.
173   vector<int> block_size_;
174 
175   // Mapping from cameras to clusters.
176   vector<int> cluster_membership_;
177 
178   // Non-zero camera pairs from the schur complement matrix that are
179   // present in the preconditioner, sorted by row (first element of
180   // each pair), then column (second).
181   set<pair<int, int> > block_pairs_;
182 
183   // Set of cluster pairs (including self pairs (i,i)) in the
184   // preconditioner.
185   HashSet<pair<int, int> > cluster_pairs_;
186   scoped_ptr<SchurEliminatorBase> eliminator_;
187 
188   // Preconditioner matrix.
189   scoped_ptr<BlockRandomAccessSparseMatrix> m_;
190 
191   // RightMultiply is a const method for LinearOperators. It is
192   // implemented using CHOLMOD's sparse triangular matrix solve
193   // function. This however requires non-const access to the
194   // SuiteSparse context object, even though it does not result in any
195   // of the state of the preconditioner being modified.
196   SuiteSparse ss_;
197 
198   // Symbolic and numeric factorization of the preconditioner.
199   cholmod_factor* factor_;
200 
201   // Temporary vector used by RightMultiply.
202   cholmod_dense* tmp_rhs_;
203   CERES_DISALLOW_COPY_AND_ASSIGN(VisibilityBasedPreconditioner);
204 };
205 #else  // SuiteSparse
206 // If SuiteSparse is not compiled in, the preconditioner is not
207 // available.
208 class VisibilityBasedPreconditioner : public BlockSparseMatrixPreconditioner {
209  public:
VisibilityBasedPreconditioner(const CompressedRowBlockStructure & bs,const Preconditioner::Options & options)210   VisibilityBasedPreconditioner(const CompressedRowBlockStructure& bs,
211                                 const Preconditioner::Options& options) {
212     LOG(FATAL) << "Visibility based preconditioning is not available. Please "
213         "build Ceres with SuiteSparse.";
214   }
~VisibilityBasedPreconditioner()215   virtual ~VisibilityBasedPreconditioner() {}
RightMultiply(const double * x,double * y)216   virtual void RightMultiply(const double* x, double* y) const {}
LeftMultiply(const double * x,double * y)217   virtual void LeftMultiply(const double* x, double* y) const {}
num_rows()218   virtual int num_rows() const { return -1; }
num_cols()219   virtual int num_cols() const { return -1; }
220 
221  private:
UpdateImpl(const BlockSparseMatrix & A,const double * D)222   bool UpdateImpl(const BlockSparseMatrix& A, const double* D) {
223     return false;
224   }
225 };
226 #endif  // CERES_NO_SUITESPARSE
227 
228 }  // namespace internal
229 }  // namespace ceres
230 
231 #endif  // CERES_INTERNAL_VISIBILITY_BASED_PRECONDITIONER_H_
232