In the Levenberg-Marquardt algorithm, the diagonal of J^T J is used to
regularize the problem. This corresponds to an elliptical trust region
|| D step || <= r, where D = sqrt(diag(J^T J)).
This commit adds the same elliptical trust region to the dogleg
strategy. The trust region problem becomes
min. x^T H x + g^T x
s.t. || D x || <= r
By substituting y = D x, it becomes
min. y^T D^-1 H D^-1 y + g^T D^-1 y
s.t. || y || <= r
which is the traditional spherical trust region problem.
This commit changes the DoglegStrategy so that the Gauss-Newton point,
the gradient, and the Cauchy point are scaled correctly (without
modifying the Jacobian directly). Then the dogleg step is computed the
same way as before, and finally the step is rescaled to obtain
x = D^-1 y.
Change-Id: Iea25a9113ecba911b746e269bc6e6fe51cb59003
There are cases where one wishes to link Ceres against another
application or library which uses a mutex implementation with
similar ancestry to the one in Ceres. In those cases there are
problems due to macro interactions which can't be contained with
namespaces.
This further isolates the Ceres Mutex class by adding CERES_
prefix to all the macros and also working around a macro name
clash with MutexLock by renaming the MutexLock class to
CeresMutexLock.
Change-Id: I923f4427d5939823ea67d48005a90391736d7751
1. Add the ability to perturb the camera pose and the
point positions using user specified parameters.
2. Re-order the flags.
3. Minor name correction.
4. Added Box-Mueller generator to random.h
Change-Id: I2c9ce74c237f5bde9a7299cc71b205d1ca9bc742
Non-monotonic trust region algorithm based on the work of Phil Toint, as
described in
Non-monotone trust region algorithms for nonlinear
optimization subject to convex constraints.
Philippe L. Toint
Mathematical Programming 77 (1997), 69-94.
Change-Id: I199ecc644e8d1a8cb43666052aef66fb93e15569
This change adds several of the pieces needed to build Ceres on
Android. The port is incomplete, since GFlags doesn't build which
causes the tests to not build. However, simple_bundle_adjuster
builds and runs on the phone.
Thanks to Scott Ettinger for the original version of the minimal
GLog implementation which made this port possible. The Ceres logs
go to the various Android logging levels, as documented in
miniglog/glog/logging.h.
To control the Android build, this adds a new CMake build option:
-DBUILD_ANDROID=ON/OFF
However, users may not want to set this manually, and should
instead run the script found in android/build_android.sh. The
script calls the NDK to make a standalone toolchain, downloads the
android-cmake toolchain, then configures CMake to cross-compile
Ceres to Android.
Note: At time of writing, the android-cmake toolchain that the
script downloads from Google Code doesn't work with the latest
Jellybean NDK. It's possible to manually hack the toolchain file
to make it build by setting the compiler version and manually
setting the path to the STL includes. I don't yet have a patch
suitable for upstreaming to http://android-cmake.googlecode.com.
Change-Id: Icb615be203145e87413d6acd05883171a395499d
Add a function to find the (complex) roots of a polynomial
with real coefficients. The roots are extracted as the
eigenvalues of the (balanced) companion matrix. Also adds a test.
The polynomial solver will be needed in the Dogleg subspace
strategy.
Change-Id: Ia6626158819efb858522b7f4998649ca010d6688
1. Added CRSMatrix object which will store the initial
and final jacobians if requested by the user.
2. Conversion routine and test for converting a
CompressedRowSparseMatrix to CRSMatrix.
3. New Evaluator::Evaluate function to do the actual evaluation.
4. Changes to Program::StateVectorToParmeterBlocks and
Program::SetParameterBlockStatePtrstoUserStatePtrs so that
they do not try to set the state of constant parameter blocks.
5. Tests for Evaluator::Evaluate.
6. Minor cleanups in SolverImpl.
7. Minor cpplint cleanups triggered by this CL.
Change-Id: I3ac446484692f943c28f2723b719676f8c83ca3d
gradient_checking_cost_function_test.cc was using arrays without initializing
them which triggers valgrind errors, and causes failures when built on the N900.
Thanks to Sebastian Koch for reporting the error and Markus Moll for pin pointing
the cause of the error.
Change-Id: Iee498969bb5026eb302ea2990edf189e70f97800
This extends the Evaluator interface to support evaluating the
gradient in addition to the residuals and jacobian, if requested.
bool Evaluate(const double* state,
double* cost,
double* residuals,
double* gradient, <----------- NEW
SparseMatrix* jacobian) = 0;
The ProgramEvaluator is extended to support the new gradient
evaluation. This required some gymnastics around the block
evaluate preparer, which now contains a scratch evaluate preparer
for the case that no jacobian is requested but the gradient is.
Gradient evaluation is a prerequisite for the planned suite of
first order methods, including nonlinear conjugate gradient,
CG_DESCENT, L-BFGS, trust region with line search, and more.
This also considerably refactors the evaluator_test to make it
shorter and check the results for all combinations of the optional
parameters [residuals, gradient, jacobian].
Change-Id: Ic7d0fec028dc5ffebc08ee079ad04eeaf6e02582
This is a preliminary, but full, port of Ceres to Windows.
Currently all tests compile and run, with only system_test
failing to work correctly due to a path issue.
Change-Id: I4152c1588bf51ffd7f4d9401ef9759f5d28c299c
Ceres has traditionally battled with portability issues
when trying to classify floating point values as one
type or another. For example, in C99 'isnan' is a
macro. Since it is a macro, it is impossible to
override the name in other namespaces.
Instead of trying to use preprocessor hacks to work
around the issue, define our own set of camel-case
names for use internally and by Ceres clients. For
example do this:
template<typename T>
void MyFunction(T x, T y) {
if (ceres::IsNaN(x)) {
...
}
}
instead of using "isnan" or "std::isnan". Note that
while GCC and Apple GCC both import 'isnan' into
the std namespace, it is not standard until C++11
which Ceres will not require for some years.
Change-Id: Ibcc96a8bb4ba63aa67cbbc58658b2e5671cd5824
This fixes the bug introduced in a previous commit,
and adds a test to check that constant parameter
blocks work as expected.
This also refactors the Solver/SolverImpl split so
that SolverImpl is no longer a friend of Problem;
instead, Solver is. This makes it possible to
verify the invariant on parameter block states in
the unit test, and is a more symmetric design
anyway.
Bug: 51
Change-Id: Id503f5b526cfb8bc24aae3aaad2e414b14063d78
1. Termination type was using == instead of =.
2. LevenbergMarquardtStrategyTest was using an object
which was not returning the correct value.
3. DumpLinearLeastSquaresProblemToTextFile was passing an
unnecessary argument to StringAppendF.
Change-Id: Ie7598c8e3d504763c889737a0d5cec805d52bbaf
1. Test that reproduces the failure on macos.
2. Move the alignment macros from manual_constructor.h
to macros.h and rename them to prevent conflicts.
3. The inline array used by FixedArray is now aligned.
4. Jet has been modified to be eigen friendly.
Change-Id: I4563847a767a92156dabab1ab420f0cdddb8ba77
User callbacks got broken at some point due to the extra
layer of copying from Solver::Options to Minimizer::Options.
This copies the user callbacks when initializing
Minimizer::Options from Solver::Options, and adds a test to
this effect.
This also fixes a bug where the state updating callback was
not called before the user callbacks. This also adds a test
to solver_impl_test to ensure the state updating callbacks
work as expected.
Thanks to Luis Alberto Zarrabeitia for the report.
Issue: 46
Change-Id: I2b36415c89dafaa5c84ecaa727a325df122e1092
The SolverImpl::Solve() method incorrectly assumed that the
state pointers inside the parameter blocks always pointed to
the user state at the start of the method. That is not true.
Change-Id: I73f8eeda453422c99e09d71a3cd0bfa92dd45742
1. Document the use of dogleg and a general discussion of
trust region methods.
2. Added a TBD section on compiler/linker flags.
3. Summary::FullReport now prints out sparse_linear_algebra_library
and trust_region_strategy_type.
Change-Id: I01f680070d510715900f345364855689005d54bb
There was a overzealous DCHECK in suitesparse.cc when converting
a scalar matrix into a block matrix. This stemmed from my poor
understanding of how lower_bound works.
The test for this function was not stringent enough, and was
not run in debug mode for this to get triggered. The test
has been updated, it fails without the fix and runs correctly
with it.
Thanks to Markus Moll for reporting this and suggesting the
fix.
Change-Id: Ide6b971fd4c618ef5e240f500f514c4b78d7b6e3
Also a fix for a minor segfault in trust_region_minimizer.cc
which was discovered while writing this test.
Change-Id: I50353d0292fd37495bf73de3824c430912ef221d
1. A new dogleg trust region strategy.
2. Consistent naming of all variables taking and reporting
time. Also all are doubles now.
3. Enum to stringification routines.
4. bundle_adjuster.cc accepts max solver time and trust_region_strategy.
5. Time accounting is pushed into solver_impl.cc and there is now
postprocessing time accounted for explicitly.
6. IterationCallback now has cumulative time.
7. LoggingCallback logs per iteration and cumulative time.
8. TrustRegionStrategy now allows for Invalid steps to be indicated
explicitly.
9. Trust region minimizer actually terminates on max_solver_time.
Change-Id: I7e3b82c8beebc17b6b355ea46ddd280754a2d8b2
SPARSE_NORMAL_CHOLESKY with CXSparse on the fairly small
bundle adjustment problem used in system_test is too expensive
to be useful as a test. It takes up too much memory and time
making the test fail on client computers < 4G of RAM.
These tests have been deleted.
Change-Id: Id015671536afd7013f5b7d19c39d64c2748884ad
Fix broken build and verbosity issues.
1. While cleaning up the last CL, I broke a macro.
2. cholmod_common_print was being called too often.
Change-Id: Ia76d8863c72f31b0c02977094b22035ceef835cf
By virtue of the modeling layer in Ceres being block oriented,
all the matrices used by Ceres are also block oriented.
When doing sparse direct factorization of these matrices, the
fill-reducing ordering algorithms can either be run on the
block or the scalar form of these matrices. Running it on the
block form exposes more of the super-nodal structure of the
matrix to the Cholesky factorization routines. This leads to
substantial gains in factorization performance.
This changelist adds support for approximate minimium degree
orderings to be computed on the block structure of the
Schur complement matrix. This affects, SchurComplementSolver
and VisibilityBasedPreconditioner and SparseNormalCholesky
when using SuiteSparse.
A bool, use_block_amd has been added to Solver::Options and
bundle_adjuster.cc has been updated to allow testing with it.
When combined with a multithreaded Schur elimination, speed ups
can be seen quite uniformly across the board. For some problems
this can be dramatic, reducing the factorization time from 70
seconds down to 17 seconds.
Change-Id: I15ebb0afcbc85ada032ec8d179ee3a2f7c8d3e46
overflow.
Even though the return value of this function is a long int, the
computation happens with three ints, which causes an overflow before
the upgrade happens.
The fix is to upgrade the constant used int his computation to be a
long int, which causes the computation to be done in longs instead of
ints.
A test has been added to verify that the fix works.
Change-Id: Ibb0aef877125bb37ca28754cb07b8e1627fd1d5a
Markus Moll