blob: 2fe025df3ce84e78dbf0cf1261657b5ccd75f7dd [file] [log] [blame]
Austin Schuh70cc9552019-01-21 19:46:48 -08001// Ceres Solver - A fast non-linear least squares minimizer
Austin Schuh1d1e6ea2020-12-23 21:56:30 -08002// Copyright 2019 Google Inc. All rights reserved.
Austin Schuh70cc9552019-01-21 19:46:48 -08003// http://ceres-solver.org/
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#ifndef CERES_PUBLIC_COVARIANCE_H_
32#define CERES_PUBLIC_COVARIANCE_H_
33
34#include <memory>
35#include <utility>
36#include <vector>
Austin Schuh1d1e6ea2020-12-23 21:56:30 -080037
Austin Schuh70cc9552019-01-21 19:46:48 -080038#include "ceres/internal/disable_warnings.h"
39#include "ceres/internal/port.h"
40#include "ceres/types.h"
41
42namespace ceres {
43
44class Problem;
45
46namespace internal {
47class CovarianceImpl;
48} // namespace internal
49
50// WARNING
51// =======
52// It is very easy to use this class incorrectly without understanding
53// the underlying mathematics. Please read and understand the
Austin Schuh1d1e6ea2020-12-23 21:56:30 -080054// documentation completely before attempting to use it.
Austin Schuh70cc9552019-01-21 19:46:48 -080055//
56//
57// This class allows the user to evaluate the covariance for a
58// non-linear least squares problem and provides random access to its
59// blocks
60//
61// Background
62// ==========
63// One way to assess the quality of the solution returned by a
64// non-linear least squares solver is to analyze the covariance of the
65// solution.
66//
67// Let us consider the non-linear regression problem
68//
69// y = f(x) + N(0, I)
70//
71// i.e., the observation y is a random non-linear function of the
72// independent variable x with mean f(x) and identity covariance. Then
73// the maximum likelihood estimate of x given observations y is the
74// solution to the non-linear least squares problem:
75//
Austin Schuh1d1e6ea2020-12-23 21:56:30 -080076// x* = arg min_x |f(x) - y|^2
Austin Schuh70cc9552019-01-21 19:46:48 -080077//
78// And the covariance of x* is given by
79//
80// C(x*) = inverse[J'(x*)J(x*)]
81//
82// Here J(x*) is the Jacobian of f at x*. The above formula assumes
83// that J(x*) has full column rank.
84//
85// If J(x*) is rank deficient, then the covariance matrix C(x*) is
86// also rank deficient and is given by
87//
88// C(x*) = pseudoinverse[J'(x*)J(x*)]
89//
90// Note that in the above, we assumed that the covariance
91// matrix for y was identity. This is an important assumption. If this
92// is not the case and we have
93//
94// y = f(x) + N(0, S)
95//
96// Where S is a positive semi-definite matrix denoting the covariance
97// of y, then the maximum likelihood problem to be solved is
98//
99// x* = arg min_x f'(x) inverse[S] f(x)
100//
101// and the corresponding covariance estimate of x* is given by
102//
103// C(x*) = inverse[J'(x*) inverse[S] J(x*)]
104//
105// So, if it is the case that the observations being fitted to have a
106// covariance matrix not equal to identity, then it is the user's
107// responsibility that the corresponding cost functions are correctly
108// scaled, e.g. in the above case the cost function for this problem
109// should evaluate S^{-1/2} f(x) instead of just f(x), where S^{-1/2}
110// is the inverse square root of the covariance matrix S.
111//
112// This class allows the user to evaluate the covariance for a
113// non-linear least squares problem and provides random access to its
114// blocks. The computation assumes that the CostFunctions compute
115// residuals such that their covariance is identity.
116//
117// Since the computation of the covariance matrix requires computing
118// the inverse of a potentially large matrix, this can involve a
119// rather large amount of time and memory. However, it is usually the
120// case that the user is only interested in a small part of the
121// covariance matrix. Quite often just the block diagonal. This class
122// allows the user to specify the parts of the covariance matrix that
123// she is interested in and then uses this information to only compute
124// and store those parts of the covariance matrix.
125//
126// Rank of the Jacobian
127// --------------------
128// As we noted above, if the jacobian is rank deficient, then the
129// inverse of J'J is not defined and instead a pseudo inverse needs to
130// be computed.
131//
132// The rank deficiency in J can be structural -- columns which are
133// always known to be zero or numerical -- depending on the exact
134// values in the Jacobian.
135//
136// Structural rank deficiency occurs when the problem contains
137// parameter blocks that are constant. This class correctly handles
138// structural rank deficiency like that.
139//
140// Numerical rank deficiency, where the rank of the matrix cannot be
141// predicted by its sparsity structure and requires looking at its
142// numerical values is more complicated. Here again there are two
143// cases.
144//
145// a. The rank deficiency arises from overparameterization. e.g., a
146// four dimensional quaternion used to parameterize SO(3), which is
147// a three dimensional manifold. In cases like this, the user should
148// use an appropriate LocalParameterization. Not only will this lead
149// to better numerical behaviour of the Solver, it will also expose
150// the rank deficiency to the Covariance object so that it can
151// handle it correctly.
152//
153// b. More general numerical rank deficiency in the Jacobian
154// requires the computation of the so called Singular Value
155// Decomposition (SVD) of J'J. We do not know how to do this for
156// large sparse matrices efficiently. For small and moderate sized
157// problems this is done using dense linear algebra.
158//
159// Gauge Invariance
160// ----------------
161// In structure from motion (3D reconstruction) problems, the
162// reconstruction is ambiguous up to a similarity transform. This is
163// known as a Gauge Ambiguity. Handling Gauges correctly requires the
164// use of SVD or custom inversion algorithms. For small problems the
165// user can use the dense algorithm. For more details see
166//
167// Ken-ichi Kanatani, Daniel D. Morris: Gauges and gauge
168// transformations for uncertainty description of geometric structure
169// with indeterminacy. IEEE Transactions on Information Theory 47(5):
170// 2017-2028 (2001)
171//
172// Example Usage
173// =============
174//
175// double x[3];
176// double y[2];
177//
178// Problem problem;
179// problem.AddParameterBlock(x, 3);
180// problem.AddParameterBlock(y, 2);
181// <Build Problem>
182// <Solve Problem>
183//
184// Covariance::Options options;
185// Covariance covariance(options);
186//
187// std::vector<std::pair<const double*, const double*>> covariance_blocks;
188// covariance_blocks.push_back(make_pair(x, x));
189// covariance_blocks.push_back(make_pair(y, y));
190// covariance_blocks.push_back(make_pair(x, y));
191//
192// CHECK(covariance.Compute(covariance_blocks, &problem));
193//
194// double covariance_xx[3 * 3];
195// double covariance_yy[2 * 2];
196// double covariance_xy[3 * 2];
197// covariance.GetCovarianceBlock(x, x, covariance_xx)
198// covariance.GetCovarianceBlock(y, y, covariance_yy)
199// covariance.GetCovarianceBlock(x, y, covariance_xy)
200//
201class CERES_EXPORT Covariance {
202 public:
203 struct CERES_EXPORT Options {
204 // Sparse linear algebra library to use when a sparse matrix
205 // factorization is being used to compute the covariance matrix.
206 //
207 // Currently this only applies to SPARSE_QR.
208 SparseLinearAlgebraLibraryType sparse_linear_algebra_library_type =
209#if !defined(CERES_NO_SUITESPARSE)
210 SUITE_SPARSE;
211#else
212 // Eigen's QR factorization is always available.
213 EIGEN_SPARSE;
214#endif
215
216 // Ceres supports two different algorithms for covariance
217 // estimation, which represent different tradeoffs in speed,
218 // accuracy and reliability.
219 //
220 // 1. DENSE_SVD uses Eigen's JacobiSVD to perform the
221 // computations. It computes the singular value decomposition
222 //
Austin Schuh1d1e6ea2020-12-23 21:56:30 -0800223 // U * D * V' = J
Austin Schuh70cc9552019-01-21 19:46:48 -0800224 //
225 // and then uses it to compute the pseudo inverse of J'J as
226 //
Austin Schuh1d1e6ea2020-12-23 21:56:30 -0800227 // pseudoinverse[J'J] = V * pseudoinverse[D^2] * V'
Austin Schuh70cc9552019-01-21 19:46:48 -0800228 //
229 // It is an accurate but slow method and should only be used
230 // for small to moderate sized problems. It can handle
231 // full-rank as well as rank deficient Jacobians.
232 //
233 // 2. SPARSE_QR uses the sparse QR factorization algorithm
234 // to compute the decomposition
235 //
236 // Q * R = J
237 //
Austin Schuh1d1e6ea2020-12-23 21:56:30 -0800238 // [J'J]^-1 = [R'*R]^-1
Austin Schuh70cc9552019-01-21 19:46:48 -0800239 //
240 // SPARSE_QR is not capable of computing the covariance if the
241 // Jacobian is rank deficient. Depending on the value of
242 // Covariance::Options::sparse_linear_algebra_library_type, either
243 // Eigen's Sparse QR factorization algorithm will be used or
244 // SuiteSparse's high performance SuiteSparseQR algorithm will be
245 // used.
246 CovarianceAlgorithmType algorithm_type = SPARSE_QR;
247
248 // If the Jacobian matrix is near singular, then inverting J'J
249 // will result in unreliable results, e.g, if
250 //
251 // J = [1.0 1.0 ]
252 // [1.0 1.0000001 ]
253 //
254 // which is essentially a rank deficient matrix, we have
255 //
256 // inv(J'J) = [ 2.0471e+14 -2.0471e+14]
257 // [-2.0471e+14 2.0471e+14]
258 //
259 // This is not a useful result. Therefore, by default
260 // Covariance::Compute will return false if a rank deficient
261 // Jacobian is encountered. How rank deficiency is detected
262 // depends on the algorithm being used.
263 //
264 // 1. DENSE_SVD
265 //
266 // min_sigma / max_sigma < sqrt(min_reciprocal_condition_number)
267 //
268 // where min_sigma and max_sigma are the minimum and maxiumum
269 // singular values of J respectively.
270 //
271 // 2. SPARSE_QR
272 //
273 // rank(J) < num_col(J)
274 //
275 // Here rank(J) is the estimate of the rank of J returned by the
276 // sparse QR factorization algorithm. It is a fairly reliable
277 // indication of rank deficiency.
278 //
279 double min_reciprocal_condition_number = 1e-14;
280
281 // When using DENSE_SVD, the user has more control in dealing with
282 // singular and near singular covariance matrices.
283 //
284 // As mentioned above, when the covariance matrix is near
285 // singular, instead of computing the inverse of J'J, the
286 // Moore-Penrose pseudoinverse of J'J should be computed.
287 //
288 // If J'J has the eigen decomposition (lambda_i, e_i), where
289 // lambda_i is the i^th eigenvalue and e_i is the corresponding
290 // eigenvector, then the inverse of J'J is
291 //
292 // inverse[J'J] = sum_i e_i e_i' / lambda_i
293 //
294 // and computing the pseudo inverse involves dropping terms from
295 // this sum that correspond to small eigenvalues.
296 //
297 // How terms are dropped is controlled by
298 // min_reciprocal_condition_number and null_space_rank.
299 //
300 // If null_space_rank is non-negative, then the smallest
301 // null_space_rank eigenvalue/eigenvectors are dropped
302 // irrespective of the magnitude of lambda_i. If the ratio of the
303 // smallest non-zero eigenvalue to the largest eigenvalue in the
304 // truncated matrix is still below
305 // min_reciprocal_condition_number, then the Covariance::Compute()
306 // will fail and return false.
307 //
308 // Setting null_space_rank = -1 drops all terms for which
309 //
310 // lambda_i / lambda_max < min_reciprocal_condition_number.
311 //
312 // This option has no effect on the SUITE_SPARSE_QR and
313 // EIGEN_SPARSE_QR algorithms.
314 int null_space_rank = 0;
315
316 int num_threads = 1;
317
318 // Even though the residual blocks in the problem may contain loss
319 // functions, setting apply_loss_function to false will turn off
320 // the application of the loss function to the output of the cost
321 // function and in turn its effect on the covariance.
322 //
323 // TODO(sameergaarwal): Expand this based on Jim's experiments.
324 bool apply_loss_function = true;
325 };
326
327 explicit Covariance(const Options& options);
328 ~Covariance();
329
330 // Compute a part of the covariance matrix.
331 //
332 // The vector covariance_blocks, indexes into the covariance matrix
333 // block-wise using pairs of parameter blocks. This allows the
334 // covariance estimation algorithm to only compute and store these
335 // blocks.
336 //
337 // Since the covariance matrix is symmetric, if the user passes
338 // (block1, block2), then GetCovarianceBlock can be called with
339 // block1, block2 as well as block2, block1.
340 //
341 // covariance_blocks cannot contain duplicates. Bad things will
342 // happen if they do.
343 //
344 // Note that the list of covariance_blocks is only used to determine
345 // what parts of the covariance matrix are computed. The full
346 // Jacobian is used to do the computation, i.e. they do not have an
347 // impact on what part of the Jacobian is used for computation.
348 //
349 // The return value indicates the success or failure of the
350 // covariance computation. Please see the documentation for
351 // Covariance::Options for more on the conditions under which this
352 // function returns false.
Austin Schuh1d1e6ea2020-12-23 21:56:30 -0800353 bool Compute(const std::vector<std::pair<const double*, const double*>>&
354 covariance_blocks,
355 Problem* problem);
Austin Schuh70cc9552019-01-21 19:46:48 -0800356
357 // Compute a part of the covariance matrix.
358 //
359 // The vector parameter_blocks contains the parameter blocks that
360 // are used for computing the covariance matrix. From this vector
361 // all covariance pairs are generated. This allows the covariance
362 // estimation algorithm to only compute and store these blocks.
363 //
364 // parameter_blocks cannot contain duplicates. Bad things will
365 // happen if they do.
366 //
367 // Note that the list of covariance_blocks is only used to determine
368 // what parts of the covariance matrix are computed. The full
369 // Jacobian is used to do the computation, i.e. they do not have an
370 // impact on what part of the Jacobian is used for computation.
371 //
372 // The return value indicates the success or failure of the
373 // covariance computation. Please see the documentation for
374 // Covariance::Options for more on the conditions under which this
375 // function returns false.
376 bool Compute(const std::vector<const double*>& parameter_blocks,
377 Problem* problem);
378
379 // Return the block of the cross-covariance matrix corresponding to
380 // parameter_block1 and parameter_block2.
381 //
382 // Compute must be called before the first call to
383 // GetCovarianceBlock and the pair <parameter_block1,
384 // parameter_block2> OR the pair <parameter_block2,
385 // parameter_block1> must have been present in the vector
386 // covariance_blocks when Compute was called. Otherwise
387 // GetCovarianceBlock will return false.
388 //
389 // covariance_block must point to a memory location that can store a
390 // parameter_block1_size x parameter_block2_size matrix. The
391 // returned covariance will be a row-major matrix.
392 bool GetCovarianceBlock(const double* parameter_block1,
393 const double* parameter_block2,
394 double* covariance_block) const;
395
396 // Return the block of the cross-covariance matrix corresponding to
397 // parameter_block1 and parameter_block2.
398 // Returns cross-covariance in the tangent space if a local
399 // parameterization is associated with either parameter block;
400 // else returns cross-covariance in the ambient space.
401 //
402 // Compute must be called before the first call to
403 // GetCovarianceBlock and the pair <parameter_block1,
404 // parameter_block2> OR the pair <parameter_block2,
405 // parameter_block1> must have been present in the vector
406 // covariance_blocks when Compute was called. Otherwise
407 // GetCovarianceBlock will return false.
408 //
409 // covariance_block must point to a memory location that can store a
410 // parameter_block1_local_size x parameter_block2_local_size matrix. The
411 // returned covariance will be a row-major matrix.
412 bool GetCovarianceBlockInTangentSpace(const double* parameter_block1,
413 const double* parameter_block2,
414 double* covariance_block) const;
415
416 // Return the covariance matrix corresponding to all parameter_blocks.
417 //
418 // Compute must be called before calling GetCovarianceMatrix and all
419 // parameter_blocks must have been present in the vector
420 // parameter_blocks when Compute was called. Otherwise
421 // GetCovarianceMatrix returns false.
422 //
423 // covariance_matrix must point to a memory location that can store
424 // the size of the covariance matrix. The covariance matrix will be
425 // a square matrix whose row and column count is equal to the sum of
426 // the sizes of the individual parameter blocks. The covariance
427 // matrix will be a row-major matrix.
Austin Schuh1d1e6ea2020-12-23 21:56:30 -0800428 bool GetCovarianceMatrix(const std::vector<const double*>& parameter_blocks,
429 double* covariance_matrix) const;
Austin Schuh70cc9552019-01-21 19:46:48 -0800430
431 // Return the covariance matrix corresponding to parameter_blocks
432 // in the tangent space if a local parameterization is associated
433 // with one of the parameter blocks else returns the covariance
434 // matrix in the ambient space.
435 //
436 // Compute must be called before calling GetCovarianceMatrix and all
437 // parameter_blocks must have been present in the vector
438 // parameters_blocks when Compute was called. Otherwise
439 // GetCovarianceMatrix returns false.
440 //
441 // covariance_matrix must point to a memory location that can store
442 // the size of the covariance matrix. The covariance matrix will be
443 // a square matrix whose row and column count is equal to the sum of
444 // the sizes of the tangent spaces of the individual parameter
445 // blocks. The covariance matrix will be a row-major matrix.
446 bool GetCovarianceMatrixInTangentSpace(
447 const std::vector<const double*>& parameter_blocks,
Austin Schuh1d1e6ea2020-12-23 21:56:30 -0800448 double* covariance_matrix) const;
Austin Schuh70cc9552019-01-21 19:46:48 -0800449
450 private:
451 std::unique_ptr<internal::CovarianceImpl> impl_;
452};
453
454} // namespace ceres
455
456#include "ceres/internal/reenable_warnings.h"
457
458#endif // CERES_PUBLIC_COVARIANCE_H_