1 // Ceres Solver - A fast non-linear least squares minimizer
2 // Copyright 2013 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 #include "ceres/covariance_impl.h"
32
33 #ifdef CERES_USE_OPENMP
34 #include <omp.h>
35 #endif
36
37 #include <algorithm>
38 #include <cstdlib>
39 #include <utility>
40 #include <vector>
41 #include "Eigen/SparseCore"
42
43 // Suppress unused local variable warning from Eigen Ordering.h #included by
44 // SparseQR in Eigen 3.2.0. This was fixed in Eigen 3.2.1, but 3.2.0 is still
45 // widely used (Ubuntu 14.04), and Ceres won't compile otherwise due to -Werror.
46 #if defined(_MSC_VER)
47 #pragma warning( push )
48 #pragma warning( disable : 4189 )
49 #else
50 #pragma GCC diagnostic push
51 #pragma GCC diagnostic ignored "-Wunused-but-set-variable"
52 #endif
53 #include "Eigen/SparseQR"
54 #if defined(_MSC_VER)
55 #pragma warning( pop )
56 #else
57 #pragma GCC diagnostic pop
58 #endif
59
60 #include "Eigen/SVD"
61 #include "ceres/compressed_col_sparse_matrix_utils.h"
62 #include "ceres/compressed_row_sparse_matrix.h"
63 #include "ceres/covariance.h"
64 #include "ceres/crs_matrix.h"
65 #include "ceres/internal/eigen.h"
66 #include "ceres/map_util.h"
67 #include "ceres/parameter_block.h"
68 #include "ceres/problem_impl.h"
69 #include "ceres/suitesparse.h"
70 #include "ceres/wall_time.h"
71 #include "glog/logging.h"
72
73 namespace ceres {
74 namespace internal {
75
76 typedef vector<pair<const double*, const double*> > CovarianceBlocks;
77
CovarianceImpl(const Covariance::Options & options)78 CovarianceImpl::CovarianceImpl(const Covariance::Options& options)
79 : options_(options),
80 is_computed_(false),
81 is_valid_(false) {
82 evaluate_options_.num_threads = options.num_threads;
83 evaluate_options_.apply_loss_function = options.apply_loss_function;
84 }
85
~CovarianceImpl()86 CovarianceImpl::~CovarianceImpl() {
87 }
88
Compute(const CovarianceBlocks & covariance_blocks,ProblemImpl * problem)89 bool CovarianceImpl::Compute(const CovarianceBlocks& covariance_blocks,
90 ProblemImpl* problem) {
91 problem_ = problem;
92 parameter_block_to_row_index_.clear();
93 covariance_matrix_.reset(NULL);
94 is_valid_ = (ComputeCovarianceSparsity(covariance_blocks, problem) &&
95 ComputeCovarianceValues());
96 is_computed_ = true;
97 return is_valid_;
98 }
99
GetCovarianceBlock(const double * original_parameter_block1,const double * original_parameter_block2,double * covariance_block) const100 bool CovarianceImpl::GetCovarianceBlock(const double* original_parameter_block1,
101 const double* original_parameter_block2,
102 double* covariance_block) const {
103 CHECK(is_computed_)
104 << "Covariance::GetCovarianceBlock called before Covariance::Compute";
105 CHECK(is_valid_)
106 << "Covariance::GetCovarianceBlock called when Covariance::Compute "
107 << "returned false.";
108
109 // If either of the two parameter blocks is constant, then the
110 // covariance block is also zero.
111 if (constant_parameter_blocks_.count(original_parameter_block1) > 0 ||
112 constant_parameter_blocks_.count(original_parameter_block2) > 0) {
113 const ProblemImpl::ParameterMap& parameter_map = problem_->parameter_map();
114 ParameterBlock* block1 =
115 FindOrDie(parameter_map,
116 const_cast<double*>(original_parameter_block1));
117
118 ParameterBlock* block2 =
119 FindOrDie(parameter_map,
120 const_cast<double*>(original_parameter_block2));
121 const int block1_size = block1->Size();
122 const int block2_size = block2->Size();
123 MatrixRef(covariance_block, block1_size, block2_size).setZero();
124 return true;
125 }
126
127 const double* parameter_block1 = original_parameter_block1;
128 const double* parameter_block2 = original_parameter_block2;
129 const bool transpose = parameter_block1 > parameter_block2;
130 if (transpose) {
131 std::swap(parameter_block1, parameter_block2);
132 }
133
134 // Find where in the covariance matrix the block is located.
135 const int row_begin =
136 FindOrDie(parameter_block_to_row_index_, parameter_block1);
137 const int col_begin =
138 FindOrDie(parameter_block_to_row_index_, parameter_block2);
139 const int* rows = covariance_matrix_->rows();
140 const int* cols = covariance_matrix_->cols();
141 const int row_size = rows[row_begin + 1] - rows[row_begin];
142 const int* cols_begin = cols + rows[row_begin];
143
144 // The only part that requires work is walking the compressed column
145 // vector to determine where the set of columns correspnding to the
146 // covariance block begin.
147 int offset = 0;
148 while (cols_begin[offset] != col_begin && offset < row_size) {
149 ++offset;
150 }
151
152 if (offset == row_size) {
153 LOG(ERROR) << "Unable to find covariance block for "
154 << original_parameter_block1 << " "
155 << original_parameter_block2;
156 return false;
157 }
158
159 const ProblemImpl::ParameterMap& parameter_map = problem_->parameter_map();
160 ParameterBlock* block1 =
161 FindOrDie(parameter_map, const_cast<double*>(parameter_block1));
162 ParameterBlock* block2 =
163 FindOrDie(parameter_map, const_cast<double*>(parameter_block2));
164 const LocalParameterization* local_param1 = block1->local_parameterization();
165 const LocalParameterization* local_param2 = block2->local_parameterization();
166 const int block1_size = block1->Size();
167 const int block1_local_size = block1->LocalSize();
168 const int block2_size = block2->Size();
169 const int block2_local_size = block2->LocalSize();
170
171 ConstMatrixRef cov(covariance_matrix_->values() + rows[row_begin],
172 block1_size,
173 row_size);
174
175 // Fast path when there are no local parameterizations.
176 if (local_param1 == NULL && local_param2 == NULL) {
177 if (transpose) {
178 MatrixRef(covariance_block, block2_size, block1_size) =
179 cov.block(0, offset, block1_size, block2_size).transpose();
180 } else {
181 MatrixRef(covariance_block, block1_size, block2_size) =
182 cov.block(0, offset, block1_size, block2_size);
183 }
184 return true;
185 }
186
187 // If local parameterizations are used then the covariance that has
188 // been computed is in the tangent space and it needs to be lifted
189 // back to the ambient space.
190 //
191 // This is given by the formula
192 //
193 // C'_12 = J_1 C_12 J_2'
194 //
195 // Where C_12 is the local tangent space covariance for parameter
196 // blocks 1 and 2. J_1 and J_2 are respectively the local to global
197 // jacobians for parameter blocks 1 and 2.
198 //
199 // See Result 5.11 on page 142 of Hartley & Zisserman (2nd Edition)
200 // for a proof.
201 //
202 // TODO(sameeragarwal): Add caching of local parameterization, so
203 // that they are computed just once per parameter block.
204 Matrix block1_jacobian(block1_size, block1_local_size);
205 if (local_param1 == NULL) {
206 block1_jacobian.setIdentity();
207 } else {
208 local_param1->ComputeJacobian(parameter_block1, block1_jacobian.data());
209 }
210
211 Matrix block2_jacobian(block2_size, block2_local_size);
212 // Fast path if the user is requesting a diagonal block.
213 if (parameter_block1 == parameter_block2) {
214 block2_jacobian = block1_jacobian;
215 } else {
216 if (local_param2 == NULL) {
217 block2_jacobian.setIdentity();
218 } else {
219 local_param2->ComputeJacobian(parameter_block2, block2_jacobian.data());
220 }
221 }
222
223 if (transpose) {
224 MatrixRef(covariance_block, block2_size, block1_size) =
225 block2_jacobian *
226 cov.block(0, offset, block1_local_size, block2_local_size).transpose() *
227 block1_jacobian.transpose();
228 } else {
229 MatrixRef(covariance_block, block1_size, block2_size) =
230 block1_jacobian *
231 cov.block(0, offset, block1_local_size, block2_local_size) *
232 block2_jacobian.transpose();
233 }
234
235 return true;
236 }
237
238 // Determine the sparsity pattern of the covariance matrix based on
239 // the block pairs requested by the user.
ComputeCovarianceSparsity(const CovarianceBlocks & original_covariance_blocks,ProblemImpl * problem)240 bool CovarianceImpl::ComputeCovarianceSparsity(
241 const CovarianceBlocks& original_covariance_blocks,
242 ProblemImpl* problem) {
243 EventLogger event_logger("CovarianceImpl::ComputeCovarianceSparsity");
244
245 // Determine an ordering for the parameter block, by sorting the
246 // parameter blocks by their pointers.
247 vector<double*> all_parameter_blocks;
248 problem->GetParameterBlocks(&all_parameter_blocks);
249 const ProblemImpl::ParameterMap& parameter_map = problem->parameter_map();
250 constant_parameter_blocks_.clear();
251 vector<double*>& active_parameter_blocks = evaluate_options_.parameter_blocks;
252 active_parameter_blocks.clear();
253 for (int i = 0; i < all_parameter_blocks.size(); ++i) {
254 double* parameter_block = all_parameter_blocks[i];
255
256 ParameterBlock* block = FindOrDie(parameter_map, parameter_block);
257 if (block->IsConstant()) {
258 constant_parameter_blocks_.insert(parameter_block);
259 } else {
260 active_parameter_blocks.push_back(parameter_block);
261 }
262 }
263
264 sort(active_parameter_blocks.begin(), active_parameter_blocks.end());
265
266 // Compute the number of rows. Map each parameter block to the
267 // first row corresponding to it in the covariance matrix using the
268 // ordering of parameter blocks just constructed.
269 int num_rows = 0;
270 parameter_block_to_row_index_.clear();
271 for (int i = 0; i < active_parameter_blocks.size(); ++i) {
272 double* parameter_block = active_parameter_blocks[i];
273 const int parameter_block_size =
274 problem->ParameterBlockLocalSize(parameter_block);
275 parameter_block_to_row_index_[parameter_block] = num_rows;
276 num_rows += parameter_block_size;
277 }
278
279 // Compute the number of non-zeros in the covariance matrix. Along
280 // the way flip any covariance blocks which are in the lower
281 // triangular part of the matrix.
282 int num_nonzeros = 0;
283 CovarianceBlocks covariance_blocks;
284 for (int i = 0; i < original_covariance_blocks.size(); ++i) {
285 const pair<const double*, const double*>& block_pair =
286 original_covariance_blocks[i];
287 if (constant_parameter_blocks_.count(block_pair.first) > 0 ||
288 constant_parameter_blocks_.count(block_pair.second) > 0) {
289 continue;
290 }
291
292 int index1 = FindOrDie(parameter_block_to_row_index_, block_pair.first);
293 int index2 = FindOrDie(parameter_block_to_row_index_, block_pair.second);
294 const int size1 = problem->ParameterBlockLocalSize(block_pair.first);
295 const int size2 = problem->ParameterBlockLocalSize(block_pair.second);
296 num_nonzeros += size1 * size2;
297
298 // Make sure we are constructing a block upper triangular matrix.
299 if (index1 > index2) {
300 covariance_blocks.push_back(make_pair(block_pair.second,
301 block_pair.first));
302 } else {
303 covariance_blocks.push_back(block_pair);
304 }
305 }
306
307 if (covariance_blocks.size() == 0) {
308 VLOG(2) << "No non-zero covariance blocks found";
309 covariance_matrix_.reset(NULL);
310 return true;
311 }
312
313 // Sort the block pairs. As a consequence we get the covariance
314 // blocks as they will occur in the CompressedRowSparseMatrix that
315 // will store the covariance.
316 sort(covariance_blocks.begin(), covariance_blocks.end());
317
318 // Fill the sparsity pattern of the covariance matrix.
319 covariance_matrix_.reset(
320 new CompressedRowSparseMatrix(num_rows, num_rows, num_nonzeros));
321
322 int* rows = covariance_matrix_->mutable_rows();
323 int* cols = covariance_matrix_->mutable_cols();
324
325 // Iterate over parameter blocks and in turn over the rows of the
326 // covariance matrix. For each parameter block, look in the upper
327 // triangular part of the covariance matrix to see if there are any
328 // blocks requested by the user. If this is the case then fill out a
329 // set of compressed rows corresponding to this parameter block.
330 //
331 // The key thing that makes this loop work is the fact that the
332 // row/columns of the covariance matrix are ordered by the pointer
333 // values of the parameter blocks. Thus iterating over the keys of
334 // parameter_block_to_row_index_ corresponds to iterating over the
335 // rows of the covariance matrix in order.
336 int i = 0; // index into covariance_blocks.
337 int cursor = 0; // index into the covariance matrix.
338 for (map<const double*, int>::const_iterator it =
339 parameter_block_to_row_index_.begin();
340 it != parameter_block_to_row_index_.end();
341 ++it) {
342 const double* row_block = it->first;
343 const int row_block_size = problem->ParameterBlockLocalSize(row_block);
344 int row_begin = it->second;
345
346 // Iterate over the covariance blocks contained in this row block
347 // and count the number of columns in this row block.
348 int num_col_blocks = 0;
349 int num_columns = 0;
350 for (int j = i; j < covariance_blocks.size(); ++j, ++num_col_blocks) {
351 const pair<const double*, const double*>& block_pair =
352 covariance_blocks[j];
353 if (block_pair.first != row_block) {
354 break;
355 }
356 num_columns += problem->ParameterBlockLocalSize(block_pair.second);
357 }
358
359 // Fill out all the compressed rows for this parameter block.
360 for (int r = 0; r < row_block_size; ++r) {
361 rows[row_begin + r] = cursor;
362 for (int c = 0; c < num_col_blocks; ++c) {
363 const double* col_block = covariance_blocks[i + c].second;
364 const int col_block_size = problem->ParameterBlockLocalSize(col_block);
365 int col_begin = FindOrDie(parameter_block_to_row_index_, col_block);
366 for (int k = 0; k < col_block_size; ++k) {
367 cols[cursor++] = col_begin++;
368 }
369 }
370 }
371
372 i+= num_col_blocks;
373 }
374
375 rows[num_rows] = cursor;
376 return true;
377 }
378
ComputeCovarianceValues()379 bool CovarianceImpl::ComputeCovarianceValues() {
380 switch (options_.algorithm_type) {
381 case DENSE_SVD:
382 return ComputeCovarianceValuesUsingDenseSVD();
383 #ifndef CERES_NO_SUITESPARSE
384 case SUITE_SPARSE_QR:
385 return ComputeCovarianceValuesUsingSuiteSparseQR();
386 #else
387 LOG(ERROR) << "SuiteSparse is required to use the "
388 << "SUITE_SPARSE_QR algorithm.";
389 return false;
390 #endif
391 case EIGEN_SPARSE_QR:
392 return ComputeCovarianceValuesUsingEigenSparseQR();
393 default:
394 LOG(ERROR) << "Unsupported covariance estimation algorithm type: "
395 << CovarianceAlgorithmTypeToString(options_.algorithm_type);
396 return false;
397 }
398 return false;
399 }
400
ComputeCovarianceValuesUsingSuiteSparseQR()401 bool CovarianceImpl::ComputeCovarianceValuesUsingSuiteSparseQR() {
402 EventLogger event_logger(
403 "CovarianceImpl::ComputeCovarianceValuesUsingSparseQR");
404
405 #ifndef CERES_NO_SUITESPARSE
406 if (covariance_matrix_.get() == NULL) {
407 // Nothing to do, all zeros covariance matrix.
408 return true;
409 }
410
411 CRSMatrix jacobian;
412 problem_->Evaluate(evaluate_options_, NULL, NULL, NULL, &jacobian);
413 event_logger.AddEvent("Evaluate");
414
415 // Construct a compressed column form of the Jacobian.
416 const int num_rows = jacobian.num_rows;
417 const int num_cols = jacobian.num_cols;
418 const int num_nonzeros = jacobian.values.size();
419
420 vector<SuiteSparse_long> transpose_rows(num_cols + 1, 0);
421 vector<SuiteSparse_long> transpose_cols(num_nonzeros, 0);
422 vector<double> transpose_values(num_nonzeros, 0);
423
424 for (int idx = 0; idx < num_nonzeros; ++idx) {
425 transpose_rows[jacobian.cols[idx] + 1] += 1;
426 }
427
428 for (int i = 1; i < transpose_rows.size(); ++i) {
429 transpose_rows[i] += transpose_rows[i - 1];
430 }
431
432 for (int r = 0; r < num_rows; ++r) {
433 for (int idx = jacobian.rows[r]; idx < jacobian.rows[r + 1]; ++idx) {
434 const int c = jacobian.cols[idx];
435 const int transpose_idx = transpose_rows[c];
436 transpose_cols[transpose_idx] = r;
437 transpose_values[transpose_idx] = jacobian.values[idx];
438 ++transpose_rows[c];
439 }
440 }
441
442 for (int i = transpose_rows.size() - 1; i > 0 ; --i) {
443 transpose_rows[i] = transpose_rows[i - 1];
444 }
445 transpose_rows[0] = 0;
446
447 cholmod_sparse cholmod_jacobian;
448 cholmod_jacobian.nrow = num_rows;
449 cholmod_jacobian.ncol = num_cols;
450 cholmod_jacobian.nzmax = num_nonzeros;
451 cholmod_jacobian.nz = NULL;
452 cholmod_jacobian.p = reinterpret_cast<void*>(&transpose_rows[0]);
453 cholmod_jacobian.i = reinterpret_cast<void*>(&transpose_cols[0]);
454 cholmod_jacobian.x = reinterpret_cast<void*>(&transpose_values[0]);
455 cholmod_jacobian.z = NULL;
456 cholmod_jacobian.stype = 0; // Matrix is not symmetric.
457 cholmod_jacobian.itype = CHOLMOD_LONG;
458 cholmod_jacobian.xtype = CHOLMOD_REAL;
459 cholmod_jacobian.dtype = CHOLMOD_DOUBLE;
460 cholmod_jacobian.sorted = 1;
461 cholmod_jacobian.packed = 1;
462
463 cholmod_common cc;
464 cholmod_l_start(&cc);
465
466 cholmod_sparse* R = NULL;
467 SuiteSparse_long* permutation = NULL;
468
469 // Compute a Q-less QR factorization of the Jacobian. Since we are
470 // only interested in inverting J'J = R'R, we do not need Q. This
471 // saves memory and gives us R as a permuted compressed column
472 // sparse matrix.
473 //
474 // TODO(sameeragarwal): Currently the symbolic factorization and the
475 // numeric factorization is done at the same time, and this does not
476 // explicitly account for the block column and row structure in the
477 // matrix. When using AMD, we have observed in the past that
478 // computing the ordering with the block matrix is significantly
479 // more efficient, both in runtime as well as the quality of
480 // ordering computed. So, it maybe worth doing that analysis
481 // separately.
482 const SuiteSparse_long rank =
483 SuiteSparseQR<double>(SPQR_ORDERING_BESTAMD,
484 SPQR_DEFAULT_TOL,
485 cholmod_jacobian.ncol,
486 &cholmod_jacobian,
487 &R,
488 &permutation,
489 &cc);
490 event_logger.AddEvent("Numeric Factorization");
491 CHECK_NOTNULL(permutation);
492 CHECK_NOTNULL(R);
493
494 if (rank < cholmod_jacobian.ncol) {
495 LOG(ERROR) << "Jacobian matrix is rank deficient. "
496 << "Number of columns: " << cholmod_jacobian.ncol
497 << " rank: " << rank;
498 free(permutation);
499 cholmod_l_free_sparse(&R, &cc);
500 cholmod_l_finish(&cc);
501 return false;
502 }
503
504 vector<int> inverse_permutation(num_cols);
505 for (SuiteSparse_long i = 0; i < num_cols; ++i) {
506 inverse_permutation[permutation[i]] = i;
507 }
508
509 const int* rows = covariance_matrix_->rows();
510 const int* cols = covariance_matrix_->cols();
511 double* values = covariance_matrix_->mutable_values();
512
513 // The following loop exploits the fact that the i^th column of A^{-1}
514 // is given by the solution to the linear system
515 //
516 // A x = e_i
517 //
518 // where e_i is a vector with e(i) = 1 and all other entries zero.
519 //
520 // Since the covariance matrix is symmetric, the i^th row and column
521 // are equal.
522 const int num_threads = options_.num_threads;
523 scoped_array<double> workspace(new double[num_threads * num_cols]);
524
525 #pragma omp parallel for num_threads(num_threads) schedule(dynamic)
526 for (int r = 0; r < num_cols; ++r) {
527 const int row_begin = rows[r];
528 const int row_end = rows[r + 1];
529 if (row_end == row_begin) {
530 continue;
531 }
532
533 # ifdef CERES_USE_OPENMP
534 int thread_id = omp_get_thread_num();
535 # else
536 int thread_id = 0;
537 # endif
538
539 double* solution = workspace.get() + thread_id * num_cols;
540 SolveRTRWithSparseRHS<SuiteSparse_long>(
541 num_cols,
542 static_cast<SuiteSparse_long*>(R->i),
543 static_cast<SuiteSparse_long*>(R->p),
544 static_cast<double*>(R->x),
545 inverse_permutation[r],
546 solution);
547 for (int idx = row_begin; idx < row_end; ++idx) {
548 const int c = cols[idx];
549 values[idx] = solution[inverse_permutation[c]];
550 }
551 }
552
553 free(permutation);
554 cholmod_l_free_sparse(&R, &cc);
555 cholmod_l_finish(&cc);
556 event_logger.AddEvent("Inversion");
557 return true;
558
559 #else // CERES_NO_SUITESPARSE
560
561 return false;
562
563 #endif // CERES_NO_SUITESPARSE
564 }
565
ComputeCovarianceValuesUsingDenseSVD()566 bool CovarianceImpl::ComputeCovarianceValuesUsingDenseSVD() {
567 EventLogger event_logger(
568 "CovarianceImpl::ComputeCovarianceValuesUsingDenseSVD");
569 if (covariance_matrix_.get() == NULL) {
570 // Nothing to do, all zeros covariance matrix.
571 return true;
572 }
573
574 CRSMatrix jacobian;
575 problem_->Evaluate(evaluate_options_, NULL, NULL, NULL, &jacobian);
576 event_logger.AddEvent("Evaluate");
577
578 Matrix dense_jacobian(jacobian.num_rows, jacobian.num_cols);
579 dense_jacobian.setZero();
580 for (int r = 0; r < jacobian.num_rows; ++r) {
581 for (int idx = jacobian.rows[r]; idx < jacobian.rows[r + 1]; ++idx) {
582 const int c = jacobian.cols[idx];
583 dense_jacobian(r, c) = jacobian.values[idx];
584 }
585 }
586 event_logger.AddEvent("ConvertToDenseMatrix");
587
588 Eigen::JacobiSVD<Matrix> svd(dense_jacobian,
589 Eigen::ComputeThinU | Eigen::ComputeThinV);
590
591 event_logger.AddEvent("SingularValueDecomposition");
592
593 const Vector singular_values = svd.singularValues();
594 const int num_singular_values = singular_values.rows();
595 Vector inverse_squared_singular_values(num_singular_values);
596 inverse_squared_singular_values.setZero();
597
598 const double max_singular_value = singular_values[0];
599 const double min_singular_value_ratio =
600 sqrt(options_.min_reciprocal_condition_number);
601
602 const bool automatic_truncation = (options_.null_space_rank < 0);
603 const int max_rank = min(num_singular_values,
604 num_singular_values - options_.null_space_rank);
605
606 // Compute the squared inverse of the singular values. Truncate the
607 // computation based on min_singular_value_ratio and
608 // null_space_rank. When either of these two quantities are active,
609 // the resulting covariance matrix is a Moore-Penrose inverse
610 // instead of a regular inverse.
611 for (int i = 0; i < max_rank; ++i) {
612 const double singular_value_ratio = singular_values[i] / max_singular_value;
613 if (singular_value_ratio < min_singular_value_ratio) {
614 // Since the singular values are in decreasing order, if
615 // automatic truncation is enabled, then from this point on
616 // all values will fail the ratio test and there is nothing to
617 // do in this loop.
618 if (automatic_truncation) {
619 break;
620 } else {
621 LOG(ERROR) << "Cholesky factorization of J'J is not reliable. "
622 << "Reciprocal condition number: "
623 << singular_value_ratio * singular_value_ratio << " "
624 << "min_reciprocal_condition_number: "
625 << options_.min_reciprocal_condition_number;
626 return false;
627 }
628 }
629
630 inverse_squared_singular_values[i] =
631 1.0 / (singular_values[i] * singular_values[i]);
632 }
633
634 Matrix dense_covariance =
635 svd.matrixV() *
636 inverse_squared_singular_values.asDiagonal() *
637 svd.matrixV().transpose();
638 event_logger.AddEvent("PseudoInverse");
639
640 const int num_rows = covariance_matrix_->num_rows();
641 const int* rows = covariance_matrix_->rows();
642 const int* cols = covariance_matrix_->cols();
643 double* values = covariance_matrix_->mutable_values();
644
645 for (int r = 0; r < num_rows; ++r) {
646 for (int idx = rows[r]; idx < rows[r + 1]; ++idx) {
647 const int c = cols[idx];
648 values[idx] = dense_covariance(r, c);
649 }
650 }
651 event_logger.AddEvent("CopyToCovarianceMatrix");
652 return true;
653 }
654
ComputeCovarianceValuesUsingEigenSparseQR()655 bool CovarianceImpl::ComputeCovarianceValuesUsingEigenSparseQR() {
656 EventLogger event_logger(
657 "CovarianceImpl::ComputeCovarianceValuesUsingEigenSparseQR");
658 if (covariance_matrix_.get() == NULL) {
659 // Nothing to do, all zeros covariance matrix.
660 return true;
661 }
662
663 CRSMatrix jacobian;
664 problem_->Evaluate(evaluate_options_, NULL, NULL, NULL, &jacobian);
665 event_logger.AddEvent("Evaluate");
666
667 typedef Eigen::SparseMatrix<double, Eigen::ColMajor> EigenSparseMatrix;
668
669 // Convert the matrix to column major order as required by SparseQR.
670 EigenSparseMatrix sparse_jacobian =
671 Eigen::MappedSparseMatrix<double, Eigen::RowMajor>(
672 jacobian.num_rows, jacobian.num_cols,
673 static_cast<int>(jacobian.values.size()),
674 jacobian.rows.data(), jacobian.cols.data(), jacobian.values.data());
675 event_logger.AddEvent("ConvertToSparseMatrix");
676
677 Eigen::SparseQR<EigenSparseMatrix, Eigen::COLAMDOrdering<int> >
678 qr_solver(sparse_jacobian);
679 event_logger.AddEvent("QRDecomposition");
680
681 if(qr_solver.info() != Eigen::Success) {
682 LOG(ERROR) << "Eigen::SparseQR decomposition failed.";
683 return false;
684 }
685
686 if (qr_solver.rank() < jacobian.num_cols) {
687 LOG(ERROR) << "Jacobian matrix is rank deficient. "
688 << "Number of columns: " << jacobian.num_cols
689 << " rank: " << qr_solver.rank();
690 return false;
691 }
692
693 const int* rows = covariance_matrix_->rows();
694 const int* cols = covariance_matrix_->cols();
695 double* values = covariance_matrix_->mutable_values();
696
697 // Compute the inverse column permutation used by QR factorization.
698 Eigen::PermutationMatrix<Eigen::Dynamic, Eigen::Dynamic> inverse_permutation =
699 qr_solver.colsPermutation().inverse();
700
701 // The following loop exploits the fact that the i^th column of A^{-1}
702 // is given by the solution to the linear system
703 //
704 // A x = e_i
705 //
706 // where e_i is a vector with e(i) = 1 and all other entries zero.
707 //
708 // Since the covariance matrix is symmetric, the i^th row and column
709 // are equal.
710 const int num_cols = jacobian.num_cols;
711 const int num_threads = options_.num_threads;
712 scoped_array<double> workspace(new double[num_threads * num_cols]);
713
714 #pragma omp parallel for num_threads(num_threads) schedule(dynamic)
715 for (int r = 0; r < num_cols; ++r) {
716 const int row_begin = rows[r];
717 const int row_end = rows[r + 1];
718 if (row_end == row_begin) {
719 continue;
720 }
721
722 # ifdef CERES_USE_OPENMP
723 int thread_id = omp_get_thread_num();
724 # else
725 int thread_id = 0;
726 # endif
727
728 double* solution = workspace.get() + thread_id * num_cols;
729 SolveRTRWithSparseRHS<int>(
730 num_cols,
731 qr_solver.matrixR().innerIndexPtr(),
732 qr_solver.matrixR().outerIndexPtr(),
733 &qr_solver.matrixR().data().value(0),
734 inverse_permutation.indices().coeff(r),
735 solution);
736
737 // Assign the values of the computed covariance using the
738 // inverse permutation used in the QR factorization.
739 for (int idx = row_begin; idx < row_end; ++idx) {
740 const int c = cols[idx];
741 values[idx] = solution[inverse_permutation.indices().coeff(c)];
742 }
743 }
744
745 event_logger.AddEvent("Inverse");
746
747 return true;
748 }
749
750 } // namespace internal
751 } // namespace ceres
752