Version History


New Features

  1. Aligned Jet matrices for improved automatic differentiation performance. (Andrew Hunter)
  2. Auto-differentiable implementations of Bessel functions, floor, and ceil (Alessandro Gentilini & Michael Vitus)
  3. New 2D and 3D SLAM examples. (Michael Vitus)
  4. Added EigenQuaternionParameterization. (Michael Vitus)
  5. Added Problem::IsParameterBlockConstant (Thomas Schneider)
  6. A complete refactoring of TrustRegionMinimizer. (Sameer Agarwal)
  7. Gradient checking cleanup and local parameterization bugfix (David Gossow)

Backward Incompatible API Changes

  1. Solver::Options::numeric_derivative_relative_step_size has been renamed to Solver::Options::gradient_check_numeric_derivative_relative_step_size. (Sameer Agarwal)

Bug Fixes & Minor Changes

  1. Clear XXX_FOUND in Find<XXX>.cmake prior to searching. (Alex Stewart)
  2. Fix versioning in the documentation (Sameer Agarwal)
  3. Fix missing gflags imported target definition in CeresConfig.cmake. (Alex Stewart)
  4. Make gflags a public dependency of Ceres if it and glog are found. (Alex Stewart)
  5. Add support for glog exported CMake target. (Alex Stewart)
  6. Use google::GLOG_WARNING instead of WARNING in tests to support MSVC. (Alex Stewart)
  7. Update gtest and gmock to a2b8a8e07628e5fd60644b6dd99c1b5e7d7f1f47 (Sameer Agarwal)
  8. Add MSVC-specific #define to expose math constants in <cmath>. (Alex Stewart)
  9. Fix typo. indepdendent -> independent (Hung Lun)
  10. Fix potential invalid reset of CMAKE_FIND_LIBRARY_PREFIXES on MSVC (Alex Stewart)
  11. Fix use of alignas(0) which is not ignored on GCC (Alex Stewart)
  12. Use default alignment if alignof(std::max_align_t) < 16 with C++11 (Alex Stewart)
  13. Introduce a common base class for DynamicAutoDiffCostFunction and DynamicNumericDiffCostFunction. (Sameer Agarwal)
  14. Fix an exact equality test causing breakage in gradient_checker_test. (Sameer Agarwal)
  15. Add GradientProblemSolver::Options::parameter_tolerance. (Sameer Agarwal)
  16. Add missing T() wrappers for constants. (Rob Carroll)
  17. Remove two checks from rotation.h (Sameer Agarwal)
  18. Relax the tolerance in QuaternionParameterizationTestHelper. (Je Hyeong Hong)
  19. Occured -> Occurred. (Sameer Agarwal)
  20. Fix a test error in (Je Hyeong Hong)
  21. Fix documentation source for templated function in rotation.h.
  22. Add package.xml to enable Catkin builds. (Damon Kohler)
  23. Relaxing Jacobian matching in Gradient Checker test. (David Gossow)
  24. Allow SubsetParameterization to hold all parameters constant (Sameer Agarwal)
  25. Fix an Intel compiler error in (Je Hyeong Hong)
  26. Removing duplicate include directive. (David Gossow)
  27. Remove two DCHECKs from CubicHermiteSpline. (Sameer Agarwal)
  28. Fix some compiler warnings. (Richard Trieu)
  29. Update ExpectArraysClose to use ExpectClose instead of EXPECT_NEAR. (Phillip Hebner)
  30. FindWithDefault returns by value rather than reference. (@aradval)
  31. Fix compiler errors on some systems. (David Gossow)
  32. Note that Problem::Evaluate cannot be called from an IterationCallback. (Sameer Agarwal)
  33. Use ProductParameterization in (Sameer Agarwal)
  34. Enable support for OpenMP in Clang if detected. (Alex Stewart)
  35. Remove duplicate entry for the NIST example in the docs. (Michael Vitus)
  36. Add additional logging for analyzing orderings (Sameer Agarwal)
  37. Add readme for the sampled_function example. (Michael Vitus)
  38. Use _j[0,1,n]() Bessel functions on MSVC to avoid deprecation errors. (Alex Stewart & Kichang Kim)
  39. Fix: Copy minimizer option is_silent to LineSearchDirection::Options (Nicolai Wojke)
  40. Fix typos in users.rst (Sameer Agarwal)
  41. Make some Jet comparisons exact. (Sameer Agarwal)
  42. Add colmap to users.rst (Sameer Agarwal)
  43. Fix step norm evaluation in LineSearchMinimizer (Sameer Agarwal)
  44. Remove use of -Werror when compiling Ceres. (Alex Stewart)
  45. Report Ceres compile options as components in find_package(). (Alex Stewart)
  46. Fix a spelling error in nnls_modeling.rst (Timer)
  47. Only use collapse() directive with OpenMP 3.0 or higher. (Keir Mierle)
  48. Fix install path for CeresConfig.cmake to be architecture-aware.
  49. Fix double conversion to degrees in rotation_test (Keir Mierle)
  50. Make Jet string output more readable (Keir Mierle)
  51. Fix rotation_test IsClose() and related tests (Keir Mierle)
  52. Loosen an exact equality in local_parameterization_test (Sameer Agarwal)
  53. make_docs: Pass the file encoding to open() (Niels Ole Salscheider)
  54. Fix error message returned when using SUITE_SPARSE_QR in covariance estimation on a ceres built without SuiteSparse support. (Simon Rutishauser)
  55. Fix CXX11 option to be available on MinGW & CygWin, but not MSVC. (Alex Stewart)
  56. Fix missing early return() in xxx_not_found() dependency macros. (Alex Stewart)
  57. Initialize inner_iterations_were_useful_ correctly. (Sameer Agarwal)
  58. Add an implementation for GradientProblemSolver::Options::IsValid (Sameer Agarwal)
  59. Fix use of va_copy() if compiling with explicit C++ version < C++11. (Alex Stewart)
  60. Install CMake files to lib/cmake/Ceres (Niels Ole Salscheider)
  61. Allow users to override the documentation install directory. (Niels Ole Salscheider)
  62. Add covariance matrix for a vector of parameters (Wannes Van Loock)
  63. Saner tolerances & stricter LRE test. (Sameer Agarwal)
  64. Fix a malformed sentence in the tutorial. (Sameer Agarwal)
  65. Add logging for sparse Cholesky factorization using Eigen. (Sameer Agarwal)
  66. Use std::adjacent_find instead of std::unique. (Sameer Agarwal)
  67. Improve logging in CompressedRowJacobianWriter on crash. (Sameer Agarwal)
  68. Fix free parameter block handling in covariance computation (Wannes Van Loock)
  69. Report the number of line search steps in FullReport. (Sameer Agarwal)
  70. Make CMake read Ceres version directly from include/ceres/version.h. (Alex Stewart)
  71. Lots of code style/lint changes. (William Rucklidge)
  72. Fix covariance computation for constant blocks (Wannes Van Loock)
  73. Add IOS_DEPLOYMENT_TARGET variable to iOS.cmake (Eduard Feicho)
  74. Make miniglog threadsafe on non-windows system by using localtime_r() instead of localtime() for time formatting (Simon Rutishauser)


New Features

  1. Adaptive numeric differentiation using Ridders’ method. (Tal Ben-Nun)
  2. Add CubicInterpolator and BiCubicInterpolator to allow smooth interpolation of sampled functions and integration with automatic differentiation.
  3. Add method to return covariance in tangent space. (Michael Vitus & Steve Hsu)
  4. Add Homogeneous vector parameterization. (Michael Vitus)
  5. Add a ProductParameterization, a local parameterization that can be constructed as a cartesian product of other local parameterization.
  6. Add DynamicCostFunctionToFunctor. (David Gossow)
  7. Optionally export Ceres build directory into local CMake package registry.
  8. Faster SPARSE_NORMAL_CHOLESKY in the presence of dynamic sparsity.

Bug Fixes & Minor Changes

  1. Remove use of link-time optimisation (LTO) for all compilers due to portability issues with gtest / type_info::operator== & Eigen with Clang on OS X vs GCC 4.9+ on Linux requiring contradictory ‘fixes’.
  2. Use link-time optimisation (LTO) only when compiling Ceres itself, not tests or examples, to bypass gtest / type_info::operator== issue.
  3. Use old minimum iOS version flags on Xcode < 7.0.
  4. Add gtest-specific flags when building/using as a shared library.
  5. Clean up iOS.cmake to use xcrun/xcodebuild & libtool.
  6. Import the latest version of googletest.
  7. Refactored system_test into bundle_adjustment_test and system_test, where each test case is its own test.
  8. Fix invalid memory access bug in CompressedRowSparseMatrix::AppendRows when it was called with a matrix of size zero.
  9. Build position independent code when compiling Ceres statically (Alexander Alekhin).
  10. Fix a bug in DetectStructure (Johannes Schonberger).
  11. Reduce memory footprint of SubsetParameterization (Johannes Schonberger).
  12. Fix for reorder program unit test when built without suitesparse (Sergey Sharybin).
  13. Fix a bug in the Schur eliminator (Werner Trobin).
  14. Fix a bug in the reordering code (Bernhard Zeisl).
  15. Add missing CERES_EXPORT to ComposedLoss (Simon Rutishauser).
  16. Add the option to use numeric differentiation to nist and more_garbow_hillstrom.
  17. Fix EIGENSPARSE option help s/t it displays in CMake ncurses GUI.
  18. Fix SparseNormalCholeskySolver with dynamic sparsity (Richie Stebbing).
  19. Remove legacy dependency detection macros.
  20. Fix failed if() condition expansion if gflags is not found.
  21. Update all CMake to lowercase function name style.
  22. Update minimum iOS version to 7.0 for shared_ptr/unordered_map.
  23. Fix bug in gflags’ <= 2.1.2 exported CMake configuration.
  24. Remove the spec file needed for generating RPMs.
  25. Fix a typo in small_blas.h (Werber Trobin).
  26. Cleanup FindGflags & use installed gflags CMake config if present.
  27. Add default glog install location on Windows to search paths (bvanevery).
  28. Add default Eigen install location on Windows to search paths (bvanevery).
  29. Fix explanation of config.h generation in bare config.h.
  30. Fix unused parameter compiler warnings in numeric_diff.h.
  31. Increase tolerance for a test in polynomial_test (Taylor Braun Jones).
  32. Fix addition of Gerrit commit hook when Ceres is a git submodule (Chris Cooper).
  33. Fix missing EIGEN_VERSION expansion typo.
  34. Fix links to SuiteSparse & CXSparse (Henrique Mendonça).
  35. Ensure Eigen is at least 3.1.0 for Eigen/SparseCore.
  36. Add option to use C++11 (not TR1) shared_ptr & unordered_map (Norman Goldstein).
  37. Fix an incorrect usage message in
  38. Gracefully disable docs if Sphinx is not found.
  39. Explicitly use (new) default OS X rpath policy if present.
  40. Add support of EIGEN_SPARSE type in IsSparseLinearAlgebraLibraryTypeAvailable function (Pierre Moulon).
  41. Allow the LossFunction contained in a LossFunctionWrapper to be NULL. This is consistent with how NULL LossFunctions are treated everywhere else. (Simon Rutishauser).
  42. Improve numeric differentation near zero.
  43. Refactored DynamicNumericDiffCostFunction to use NumericDiff (Tal Ben-Nun).
  44. Remove use of :caption tag in Sphinx.
  45. Add a small test to make sure GradientProblemSolver works correctly (Petter Strandmark).
  46. Add simple unit tests for GradientProblem (Petter Strandmark).
  47. Make the robust curve fitting example robust.
  48. Homogenize convergence operators in docs and code (Johannes Schonberger).
  49. Add parameter_tolerance convergence to line search minimizer (Johannes Schonberger).
  50. Fix bug where pow(JetA,JetB) returned wrong result for JetA==0 (Russell Smith).
  51. Remove duplicate step norm computation (Johannes Schonberger).
  52. Enhance usability when encountering Eigen version mismatches (Andrew Hundt).
  53. Add PLY file logger before and after BA in order to ease visual comparison (Pierre Moulon).
  54. Fix CMake config file docs to include 2.8.x & 3.x styles.
  55. Python3 fixes (Markus Moll).
  56. Remove confusing code from DenseJacobianWriter (Michael Vitus).
  57. Add documentation on CMake package installation process.
  58. Revert a call to SolveUpperTriangularUsingCholesky.
  59. Make CERES_EIGEN_VERSION macro independent of CMake.
  60. Add versions of dependencies used to FullReport().
  61. Ensure local config.h is used if Ceres is already installed.
  62. Small messaging and comment updates in CMake
  63. Handle possible presence of library prefixes in MSVC (Sylvain Duchêne).
  64. Use -O2 not -O3 on MinGW to workaround issue with Eigen (
  65. Increase tolerance in small_blas test for Cygwin (
  66. Fix iOS cmake file for cmake 3.0 (Jack Feng)
  67. Fix missing gflags shlwapi dependency on MinGW (
  68. Add thread dependency & fix namespace detection on Windows for gflags (
  69. Rename macros in the public API to have a CERES_ prefix.
  70. Fix OrderedGroup::Reverse() when it is empty (Chris Sweeney).
  71. Update the code to point to
  72. Update documentation to point to the GitHub issue tracker.
  73. Disable LAPACK for iOS builds. (Greg Coombe)
  74. Force use of single-thread in Problem::Evaluate() without OpenMP.
  75. Less strict check for multithreading. (Chris Sweeney)
  76. Update tolerances in (Philipp Hubner)
  77. Documentation corrections (Steve Hsu)
  78. Fixed (Pablo Speciale)
  79. Fix example code in the documentation. (Rodney Hoskinson)
  80. Improve the error handling in Conjugate Gradients.
  81. Improve preconditioner documentation.
  82. Remove dead code from fpclassify.h.
  83. Make threads sensitive.
  84. Changed the CURRENT_CONFIG_INSTALL_DIR to be a variable local to Ceres. (Chris Sweeney)
  85. Fix typo in the comments in Jet.h. (Julius Ziegler)
  86. Add the ASL at ETH Zurich, Theia & OpenPTrack to the list of users.
  87. Fixed a typo in the documentation. (Richard Stebbing)
  88. Fixed a boundary handling bug in the BiCubic interpolation code. (Bernhard Zeisl)
  89. Fixed a MSVC compilation bug in the cubic interpolation code (Johannes Schönberger)
  90. Add covariance related files to the Android build.
  91. Update Ubuntu 14.04 installation instructions. (Filippo Basso)
  92. Improved logging for linear solver failures.
  93. Improved crash messages in Problem.
  94. Hide Homebrew related variables in CMake GUI.
  95. Add SuiteSparse link dependency for compressed_col_sparse_matrix_utils_test.
  96. Autodetect Homebrew install prefix on OSX.
  97. Lint changes from William Rucklidge and Jim Roseborough.
  98. Remove using namespace std: from port.h
  99. Add note about glog not currently compiling against gflags 2.1.
  100. Add explicit no sparse linear algebra library available option.
  101. Improve some wording in the FAQ. (Vasily Vylkov)
  102. Delete Incomplete LQ Factorization.
  103. Add a pointer to MacPorts. (Markus Moll)


New Features

  1. Ceres Solver can now be used to solve general unconstrained optimization problems. See the documentation for GradientProblem and GradientProblemSolver.

  2. Eigen can now be as a sparse linear algebra backend. This can be done by setting Solver::Options::sparse_linear_algebra_library_type to EIGEN_SPARSE. Performance should be comparable to CX_SPARSE.


    Because Eigen is a header only library, and some of the code related to sparse Cholesky factorization is LGPL, building Ceres with support for Eigen’s sparse linear algebra is disabled by default and should be enabled explicitly.


    For good performance, use Eigen version 3.2.2 or later.

  3. Added EIGEN_SPARSE_QR algorithm for covariance estimation using Eigen‘s sparse QR factorization. (Michael Vitus)

  4. Faster inner iterations when using multiple threads.

  5. Faster ITERATIVE_SCHUR + SCHUR_JACOBI for small to medium sized problems (see documentation for Solver::Options::use_explicit_schur_complement).

  6. Faster automatic Schur ordering.

  7. Reduced memory usage when solving problems with dynamic sparsity.

  8. CostFunctionToFunctor now supports dynamic number of residuals.

  9. A complete re-write of the problem preprocessing phase.

  10. Solver::Summary::FullReport now reports the build configuration for Ceres.

  11. When building on Android, the NDK version detection logic has been improved.

  12. The CERES_VERSION macro has been improved and replaced with the CERES_VERSION_STRING macro.

  13. Added Solver::Options::IsValid which allows users to validate their solver configuration before calling Solve.

  14. Added Problem::GetCostFunctionForResidualBlock and Problem::GetLossFunctionForResidualBlock.

  15. Added Tukey’s loss function. (Michael Vitus)

  16. Added RotationMatrixToQuaternion

  17. Compute & report timing information for line searches.

  18. Autodetect gflags namespace.

  19. Expanded

  20. Added a pointer to Tal Ben-Nun’s MSVC wrapper to the docs.

  21. Added the <2,3,6> Schur template specialization. (Alessandro Dal Grande)

Backward Incompatible API Changes

  1. NumericDiffFunctor has been removed. It’s API was broken, and the implementation was an unnecessary layer of abstraction over CostFunctionToFunctor.
  2. POLAK_RIBIRERE conjugate gradients direction type has been renamed to POLAK_RIBIERE.
  3. Solver::Options::solver_log has been removed. If needed this iteration callback can easily be implemented in user code.
  4. The SPARSE_CHOLESKY algorithm for covariance estimation has been removed. It is not rank revealing and numerically poorly behaved. Sparse QR factorization is a much better way to do this.
  5. The SPARSE_QR algorithm for covariance estimation has been renamed to SUITE_SPARSE_QR to be consistent with EIGEN_SPARSE_QR.
  6. Solver::Summary::preconditioner_type has been replaced with Solver::Summary::preconditioner_type_given and Solver::Summary::preconditioner_type_used to be more consistent with how information about the linear solver is communicated.
  7. CERES_VERSION and CERES_ABI_VERSION macros were not terribly useful. They have been replaced with CERES_VERSION_MAJOR, CERES_VERSION_MINOR , CERES_VERSION_REVISION and CERES_VERSION_ABI macros. In particular the functionality of CERES_VERSION is provided by CERES_VERSION_STRING macro.

Bug Fixes

  1. Do not try the gradient step if TR step line search fails.
  2. Fix missing include in libmv_bundle_adjuster on OSX.
  3. Conditionally log evaluation failure warnings.
  4. Runtime uses four digits after the decimal in Summary:FullReport.
  5. Better options checking for TrustRegionMinimizer.
  6. Fix RotationMatrixToAngleAxis when the angle of rotation is near PI. (Tobias Strauss)
  7. Sometimes gradient norm based convergence would miss a step with a substantial solution quality improvement. (Rodney Hoskinson)
  8. Ignore warnings from within Eigen/SparseQR (3.2.2).
  9. Fix empty Cache HELPSTRING parsing error on OS X 10.10 Yosemite.
  10. Fix a formatting error TrustRegionMinimizer logging.
  11. Add an explicit include for local_parameterization.h (cooordz)
  12. Fix a number of typos in the documentation (Martin Baeuml)
  13. Made the logging in TrustRegionMinimizer consistent with LineSearchMinimizer.
  14. Fix some obsolete documentation in CostFunction::Evaluate.
  15. Fix CG solver options for ITERATIVE_SCHUR, which did not copy min_num_iterations (Johannes Schönberger)
  16. Remove obsolete include of numeric_diff_functor.h. (Martin Baeuml)
  17. Fix max. linear solver iterations in ConjugateGradientsSolver (Johannes Schönberger)
  18. Expand check for lack of a sparse linear algebra library. (Michael Samples and Domink Reitzle)
  19. Fix Eigen Row/ColMajor bug in NumericDiffCostFunction. (Dominik Reitzle)
  20. Fix crash in Covariance if # threads > 1 requested without OpenMP.
  21. Fixed Malformed regex. (Björn Piltz)
  22. Fixed MSVC error C2124: divide or mod by zero. (Björn Piltz)
  23. Add missing #include of <limits> for loss functions.
  24. Make canned loss functions more robust.
  25. Fix type of suppressed compiler warning for Eigen 3.2.0.
  26. Suppress unused variable warning from Eigen 3.2.0.
  27. Add “make install” to the install instructions.
  28. Correct formula in documentation of Solver::Options::function_tolerance. (Alessandro Gentilini)
  29. Add release flags to iOS toolchain.
  30. Fix a broken hyperlink in the documentation. (Henrique Mendonca)
  31. Add fixes for multiple definitions of ERROR on Windows to docs.
  32. Compile miniglog into Ceres if enabled on all platforms.
  33. Add two missing files to (Greg Coombe)
  34. Fix Cmake error when using miniglog. (Greg Coombe)
  35. Don’t build miniglog unconditionally as a static library (Björn Piltz)
  36. Added a missing include. (Björn Piltz)
  37. Conditionally disable SparseNormalCholesky.
  38. Fix a memory leak in


New Features

  1. Bounds constraints: Support for upper and/or lower bounds on parameters when using the trust region minimizer.
  2. Dynamic Sparsity: Problems in which the sparsity structure of the Jacobian changes over the course of the optimization can now be solved much more efficiently. (Richard Stebbing)
  3. Improved support for Microsoft Visual C++ including the ability to build and ship DLLs. (Björn Piltz, Alex Stewart and Sergey Sharybin)
  4. Support for building on iOS 6.0 or higher (Jack Feng).
  5. Autogeneration of config.h that captures all the defines used to build and use Ceres Solver.
  6. Simpler and more informative solver termination type reporting. (See below for more details)
  7. New website based entirely on Sphinx.
  8. AutoDiffLocalParameterization allows the use of automatic differentiation for defining LocalParameterization objects (Alex Stewart)
  9. LBFGS is faster due to fewer memory copies.
  10. Parameter blocks are not restricted to be less than 32k in size, they can be up to 2G in size.
  11. Faster SPARSE_NORMAL_CHOLESKY solver when using CX_SPARSE as the sparse linear algebra library.
  12. Added Problem::IsParameterBlockPresent and Problem::GetParameterization.
  13. Added the (2,4,9) and (2,4,8) template specializations.
  14. An example demonstrating the use of DynamicAutoDiffCostFunction. (Joydeep Biswas)
  15. Homography estimation example from Blender demonstrating the use of a custom IterationCallback. (Sergey Sharybin)
  16. Support user passing a custom CMAKE_MODULE_PATH (for BLAS / LAPACK).

Backward Incompatible API Changes

  1. Solver::Options::linear_solver_ordering used to be a naked pointer that Ceres took ownership of. This is error prone behaviour which leads to problems when copying the Solver::Options struct around. This has been replaced with a shared_ptr to handle ownership correctly across copies.

  2. The enum used for reporting the termination/convergence status of the solver has been renamed from SolverTerminationType to TerminationType.

    The enum values have also changed. FUNCTION_TOLERANCE, GRADIENT_TOLERANCE and PARAMETER_TOLERANCE have all been replaced by CONVERGENCE.

    NUMERICAL_FAILURE has been replaed by FAILURE.

    USER_ABORT has been renamed to USER_FAILURE.

    Further Solver::Summary::error has been renamed to Solver::Summary::message. It contains a more detailed explanation for why the solver terminated.

  3. Solver::Options::gradient_tolerance used to be a relative gradient tolerance. i.e., The solver converged when

    \[\|g(x)\|_\infty < \text{gradient_tolerance} * \|g(x_0)\|_\infty\]

    where \(g(x)\) is the gradient of the objective function at \(x\) and \(x_0\) is the parmeter vector at the start of the optimization.

    This has changed to an absolute tolerance, i.e. the solver converges when

    \[\|g(x)\|_\infty < \text{gradient_tolerance}\]
  4. Ceres cannot be built without the line search minimizer anymore. Thus the preprocessor define CERES_NO_LINE_SEARCH_MINIMIZER has been removed.

Bug Fixes

  1. Disabled warning C4251. (Björn Piltz)
  2. Do not propagate 3d party libs through IMPORTED_LINK_INTERFACE_LIBRARIES_[DEBUG/RELEASE] mechanism when building shared libraries. (Björn Piltz)
  3. Fixed errant verbose levels (Björn Piltz)
  4. Variety of code cleanups, optimizations and bug fixes to the line search minimizer code (Alex Stewart)
  5. Fixed BlockSparseMatrix::Transpose when the matrix has row and column blocks. (Richard Bowen)
  6. Better error checking when Problem::RemoveResidualBlock is called. (Alex Stewart)
  7. Fixed a memory leak in SchurComplementSolver.
  8. Added epsilon() method to NumTraits<ceres::Jet<T, N> >. (Filippo Basso)
  9. Fixed a bug in CompressedRowSparseMatrix::AppendRows` and DeleteRows.q
  10. Handle empty problems consistently.
  11. Restore the state of the Problem after a call to Problem::Evaluate. (Stefan Leutenegger)
  12. Better error checking and reporting for linear solvers.
  13. Use explicit formula to solve quadratic polynomials instead of the eigenvalue solver.
  14. Fix constant parameter handling in inner iterations (Mikael Persson).
  15. SuiteSparse errors do not cause a fatal crash anymore.
  16. Fix
  17. Relax the requirements on loss function derivatives.
  18. Minor bugfix to logging.h (Scott Ettinger)
  19. Updated gmock and gtest to the latest upstream version.
  20. Fix build breakage on old versions of SuiteSparse.
  21. Fixed build issues related to Clang / LLVM 3.4 (Johannes Schönberger)
  22. METIS_FOUND is never set. Changed the commit to fit the setting of the other #._FOUND definitions. (Andreas Franek)
  23. Variety of bug fixes and cleanups to the CMake build system (Alex Stewart)
  24. Removed fictious shared library target from the NDK build.
  25. Solver::Options now uses shared_ptr to handle ownership of Solver::Options::linear_solver_ordering and Solver::Options::inner_iteration_ordering. As a consequence the NDK build now depends on libc++ from the LLVM project.
  26. Variety of lint cleanups (William Rucklidge & Jim Roseborough)
  27. Various internal cleanups including dead code removal.


New Features

  1. Significant improved CMake files with better robustness, dependency checking and GUI support. (Alex Stewart)
  2. Added DynamicNumericDiffCostFunction for numerically differentiated cost functions whose sizing is determined at run time.
  3. NumericDiffCostFunction now supports a dynamic number of residuals just like AutoDiffCostFunction.
  4. Problem exposes more of its structure in its API.
  5. Faster automatic differentiation (Tim Langlois)
  6. Added the commonly occuring 2_d_d template specialization for the Schur Eliminator.
  7. Faster ITERATIVE_SCHUR solver using template specializations.
  8. Faster SCHUR_JACOBI preconditioner construction.
  9. Faster AngleAxisRotatePoint.
  10. Faster Jacobian evaluation when a loss function is used.
  11. Added support for multiple clustering algorithms in visibility based preconditioning, including a new fast single linkage clustering algorithm.

Bug Fixes

  1. Fix ordering of ParseCommandLineFlags() & InitGoogleTest() for Windows. (Alex Stewart)
  2. Remove DCHECK_GE checks from fixed_array.h.
  3. Fix build on MSVC 2013 (Petter Strandmark)
  4. Fixed AngleAxisToRotationMatrix near zero.
  5. Move CERES_HASH_NAMESPACE macros to collections_port.h.
  6. Fix handling of unordered_map/unordered_set on OSX 10.9.0.
  7. Explicitly link to libm for curve_fitting_c.c. (Alex Stewart)
  8. Minor type conversion fix to autodiff.h
  9. Remove RuntimeNumericDiffCostFunction.
  10. Fix operator= ambiguity on some versions of Clang. (Alex Stewart)
  11. Various Lint cleanups (William Rucklidge & Jim Roseborough)
  12. Modified installation folders for Windows. (Pablo Speciale)
  13. Added librt to link libraries for SuiteSparse_config on Linux. (Alex Stewart)
  14. Check for presence of return-type-c-linkage option with Clang. (Alex Stewart)
  15. Fix Problem::RemoveParameterBlock after calling solve. (Simon Lynen)
  16. Fix a free/delete bug in
  17. Fix two build errors. (Dustin Lang)
  18. Add RequireInitialization = 1 to NumTraits::Jet.
  19. Update gmock/gtest to 1.7.0
  20. Added IterationSummary::gradient_norm.
  21. Reduced verbosity of the inner iteration minimizer.
  22. Fixed a bug in TrustRegionMinimizer. (Michael Vitus)
  23. Removed android/


Backward Incompatible API Changes

  1. Solver::Options::sparse_linear_algebra_library has been renamed to Solver::Options::sparse_linear_algebra_library_type.

New Features

  1. Sparse and dense covariance estimation.
  2. A new Wolfe line search. (Alex Stewart)
  3. BFGS line search direction. (Alex Stewart)
  4. C API
  5. Speeded up the use of loss functions > 17x.
  7. Support for multiple dense linear algebra backends. In particular optimized BLAS and LAPACK implementations (e.g., Intel MKL, ACML, OpenBLAS etc) can now be used to do the dense linear algebra for DENSE_QR, DENSE_NORMAL_CHOLESKY and DENSE_SCHUR
  8. Use of Inner iterations can now be adaptively stopped. Iteration and runtime statistics for inner iterations are not reported in Solver::Summary and Solver::Summary::FullReport.
  9. Improved inner iteration step acceptance criterion.
  10. Add BlockRandomAccessCRSMatrix.
  11. Speeded up automatic differentiation by 7%.
  12. Bundle adjustment example from libmv/Blender (Sergey Sharybin)
  13. Shared library building is now controlled by CMake, rather than a custom solution. Previously, Ceres had a custom option, but this is now deprecated in favor of CMake’s built in support for switching between static and shared. Turn on BUILD_SHARED_LIBS to get shared Ceres libraries.
  14. No more dependence on Protocol Buffers.
  15. Incomplete LQ factorization.
  16. Ability to write trust region problems to disk.
  17. Add sinh, cosh, tanh and tan functions to automatic differentiation (Johannes Schönberger)
  18. Simplifications to the cmake build file.
  19. miniglog can now be used as a replacement for google-glog on non Android platforms. (This is NOT recommended).

Bug Fixes

  1. Fix ITERATIVE_SCHUR solver to work correctly when the schur complement is of size zero. (Soohyun Bae)
  2. Fix the spec file for generating RPM packages (Brian Pitts and Taylor Braun-Jones).
  3. Fix how ceres calls CAMD (Manas Jagadev)
  4. Fix breakage on old versions of SuiteSparse. (Fisher Yu)
  5. Fix warning C4373 in Visual Studio (Petter Strandmark)
  6. Fix compilation error caused by missing suitesparse headers and reorganize them to be more robust. (Sergey Sharybin)
  7. Check GCC Version before adding -fast compiler option on OSX. (Steven Lovegrove)
  8. Add documentation for minimizer progress output.
  9. Lint and other cleanups (William Rucklidge and James Roseborough)
  10. Collections port fix for MSC 2008 (Sergey Sharybin)
  11. Various corrections and cleanups in the documentation.
  12. Change the path where CeresConfig.cmake is installed (Pablo Speciale)
  13. Minor errors in documentation (Pablo Speciale)
  14. Updated depend.cmake to follow CMake IF convention. (Joydeep Biswas)
  15. Stablize the schur ordering algorithm.
  16. Update license header in split.h.
  17. Enabling -O4 (link-time optimization) only if compiler/linker support it. (Alex Stewart)
  18. Consistent glog path across files.
  19. ceres-solver.spec: Use cleaner, more conventional Release string (Taylor Braun-Jones)
  20. Fix compile bug on RHEL6 due to missing header (Taylor Braun-Jones)
  21. CMake file is less verbose.
  22. Use the latest upstream version of google-test and gmock.
  23. Rationalize some of the variable names in Solver::Options.
  24. Improve Summary::FullReport when line search is used.
  25. Expose line search parameters in Solver::Options.
  26. Fix update of L-BFGS history buffers after they become full. (Alex Stewart)
  27. Fix configuration error on systems without SuiteSparse installed (Sergey Sharybin)
  28. Enforce the read call returns correct value in curve_fitting_c.c (Arnaud Gelas)
  29. Fix DynamicAutoDiffCostFunction (Richard Stebbing)
  30. Fix Problem::RemoveParameterBlock documentation (Johannes Schönberger)
  31. Fix a logging bug in parameter_block.h
  32. Refactor the preconditioner class structure.
  33. Fix an uninitialized variable warning when building with GCC.
  34. Fix a reallocation bug in CreateJacobianBlockSparsityTranspose. (Yuliy Schwartzburg)
  35. Add a define for O_BINARY.
  36. Fix miniglog-based Android NDK build; now works with NDK r9. (Scott Ettinger)


New Features

  1. Major Performance improvements.
    1. Schur type solvers (SPARSE_SCHUR, DENSE_SCHUR, ITERATIVE_SCHUR) are significantly faster due to custom BLAS routines and fewer heap allocations.
    2. SPARSE_SCHUR when used with CX_SPARSE now uses a block AMD for much improved factorization performance.
    3. The jacobian matrix is pre-ordered so that SPARSE_NORMAL_CHOLESKY and SPARSE_SCHUR do not have to make copies inside CHOLMOD.
    4. Faster autodiff by replacing division by multplication by inverse.
    5. When compiled without threads, the schur eliminator does not pay the penalty for locking and unlocking mutexes.
  2. Users can now use linear_solver_ordering to affect the fill-reducing ordering used by SUITE_SPARSE for SPARSE_NORMAL_CHOLESKY.
  3. Problem can now report the set of parameter blocks it knows about.
  4. TrustRegionMinimizer uses the evaluator to compute the gradient instead of a matrix vector multiply.
  5. On Mac OS, whole program optimization is enabled.
  6. Users can now use automatic differentiation to define new LocalParameterization objects. (Sergey Sharybin)
  7. Enable larger tuple sizes for Visual Studio 2012. (Petter Strandmark)

Bug Fixes

  1. Update the documentation for CostFunction.
  2. Fixed a typo in the documentation. (Pablo Speciale)
  3. Fix a typo in
  4. Bugfix in NumericDiffCostFunction. (Nicolas Brodu)
  5. Death to BlockSparseMatrixBase.
  6. Change Minimizer::Options::min_trust_region_radius to double.
  7. Update to compile with stricter gcc checks. (Joydeep Biswas)
  8. Do not modify cached CMAKE_CXX_FLAGS_RELEASE. (Sergey Sharybin)
  9. <iterator> needed for back_insert_iterator. (Petter Strandmark)
  10. Lint cleanup. (William Rucklidge)
  11. Documentation corrections. (Pablo Speciale)


Backward Incompatible API Changes

  1. Added Problem::Evaluate. Now you can evaluate a problem or any part of it without calling the solver.

    In light of this the following settings have been deprecated and removed from the API.

    • Solver::Options::return_initial_residuals
    • Solver::Options::return_initial_gradient
    • Solver::Options::return_initial_jacobian
    • Solver::Options::return_final_residuals
    • Solver::Options::return_final_gradient
    • Solver::Options::return_final_jacobian

    Instead we recommend using something like this.

    Problem problem;
    // Build problem
    vector<double> initial_residuals;
                     NULL, /* No cost */
                     NULL, /* No gradient */
                     NULL  /* No jacobian */ );
    Solver::Options options;
    Solver::Summary summary;
    Solver::Solve(options, &problem, &summary);
    vector<double> final_residuals;
                     NULL, /* No cost */
                     NULL, /* No gradient */
                     NULL  /* No jacobian */ );

New Features

  1. Problem now supports removal of ParameterBlocks and ResidualBlocks. There is a space/time tradeoff in doing this which is controlled by Problem::Options::enable_fast_parameter_block_removal.
  2. Ceres now supports Line search based optimization algorithms in addition to trust region algorithms. Currently there is support for gradient descent, non-linear conjugate gradient and LBFGS search directions.
  3. Added Problem::Evaluate. Now you can evaluate a problem or any part of it without calling the solver. In light of this the following settings have been deprecated and removed from the API.
    • Solver::Options::return_initial_residuals
    • Solver::Options::return_initial_gradient
    • Solver::Options::return_initial_jacobian
    • Solver::Options::return_final_residuals
    • Solver::Options::return_final_gradient
    • Solver::Options::return_final_jacobian
  4. New, much improved HTML documentation using Sphinx.
  5. Changed NumericDiffCostFunction to take functors like AutoDiffCostFunction.
  6. Added support for mixing automatic, analytic and numeric differentiation. This is done by adding CostFunctionToFunctor and NumericDiffFunctor objects to the API.
  7. Sped up the robust loss function correction logic when residual is one dimensional.
  8. Sped up DenseQRSolver by changing the way dense jacobians are stored. This is a 200-500% improvement in linear solver performance depending on the size of the problem.
  9. DENSE_SCHUR now supports multi-threading.
  10. Greatly expanded Summary::FullReport:
    • Report the ordering used by the LinearSolver.
    • Report the ordering used by the inner iterations.
    • Execution timing breakdown into evaluations and linear solves.
    • Effective size of the problem solved by the solver, which now accounts for the size of the tangent space when using a LocalParameterization.
  11. Ceres when run at the VLOG level 3 or higher will report detailed timing information about its internals.
  12. Remove extraneous initial and final residual evaluations. This speeds up the solver a bit.
  13. Automatic differenatiation with a dynamic number of parameter blocks. (Based on an idea by Thad Hughes).
  14. Sped up problem construction and destruction.
  15. Added matrix adapters to rotation.h so that the rotation matrix routines can work with row and column major matrices. (Markus Moll)
  16. SCHUR_JACOBI can now be used without SuiteSparse.
  17. A .spec file for producing RPMs. (Taylor Braun-Jones)
  18. CMake can now build the sphinx documentation (Pablo Speciale)
  19. Add support for creating a CMake config file during build to make embedding Ceres in other CMake-using projects easier. (Pablo Speciale).
  20. Better error reporting in Problem for missing parameter blocks.
  21. A more flexible and a more modular build. If binary size and/or compile time is a concern, larger parts of the solver can be disabled at compile time.

Bug Fixes

  1. Compilation fixes for MSVC2010 (Sergey Sharybin)
  2. Fixed “deprecated conversion from string constant to char*” warnings. (Pablo Speciale)
  3. Correctly propagate ifdefs when building without Schur eliminator template specializations.
  4. Correct handling of LIB_SUFFIX on Linux. (Yuliy Schwartzburg).
  5. Code and signature cleanup in rotation.h.
  6. Make examples independent of internal code.
  7. Disable unused member in gtest which results in build error on OS X with latest Xcode. (Taylor Braun-Jones)
  8. Pass the correct flags to the linker when using pthreads. (Taylor Braun-Jones)
  9. Only use cmake28 macro when building on RHEL6. (Taylor Braun-Jones)
  10. Remove -Wno-return-type-c-linkage when compiling with GCC. (Taylor Braun-Jones)
  11. Fix No previous prototype warnings. (Sergey Sharybin)
  12. MinGW build fixes. (Sergey Sharybin)
  13. Lots of minor code and lint fixes. (William Rucklidge)
  14. Fixed a bug in residual evaluation. (Markus Moll)
  15. Fixed varidic evaluation bug in AutoDiff.
  16. Fixed SolverImpl tests.
  17. Fixed a bug in DenseSparseMatrix::ToDenseMatrix().
  18. Fixed an initialization bug in ProgramEvaluator.
  19. Fixes to paths (Carlos Hernandez)
  20. Modify to compute accuracy based on ground truth solution rather than the ground truth function value.
  21. Fixed a memory leak in (Alexander Mordvintsev).
  22. Fixed the install directory for libraries by correctly handling LIB_SUFFIX. (Taylor Braun-Jones)


Backward Incompatible API Changes

The new ordering API breaks existing code. Here the common case fixes.


options.linear_solver_type = ceres::DENSE_SCHUR
options.ordering_type = ceres::SCHUR


options.linear_solver_type = ceres::DENSE_SCHUR


options.linear_solver_type = ceres::DENSE_SCHUR;
options.ordering_type = ceres::USER;
for (int i = 0; i < num_points; ++i) {
for (int i = 0; i < num_cameras; ++i) {
options.num_eliminate_blocks = num_points;


options.linear_solver_type = ceres::DENSE_SCHUR;
options.ordering = new ceres::ParameterBlockOrdering;
for (int i = 0; i < num_points; ++i) {
  options.linear_solver_ordering->AddElementToGroup(my_points[i], 0);
for (int i = 0; i < num_cameras; ++i) {
  options.linear_solver_ordering->AddElementToGroup(my_cameras[i], 1);

New Features

  1. A new richer, more expressive and consistent API for ordering parameter blocks.
  2. A non-linear generalization of Ruhe & Wedin’s Algorithm II. This allows the user to use variable projection on separable and non-separable non-linear least squares problems. With multithreading, this results in significant improvements to the convergence behavior of the solver at a small increase in run time.
  3. An image denoising example using fields of experts. (Petter Strandmark)
  4. Defines for Ceres version and ABI version.
  5. Higher precision timer code where available. (Petter Strandmark)
  6. Example Makefile for users of Ceres.
  7. IterationSummary now informs the user when the step is a non-monotonic step.
  8. Fewer memory allocations when using DenseQRSolver.
  9. GradientChecker for testing CostFunctions (William Rucklidge)
  10. Add support for cost functions with 10 parameter blocks in Problem. (Fisher)
  11. Add support for 10 parameter blocks in AutoDiffCostFunction.

Bug Fixes

  1. static cast to force Eigen::Index to long conversion
  2. Change LOG(ERROR) to LOG(WARNING) in
  3. Remove verbose logging from DenseQRSolve.
  4. Fix the Android NDK build.
  5. Better handling of empty and constant Problems.
  6. Remove an internal header that was leaking into the public API.
  7. Memory leak in
  8. Schur ordering was operating on the wrong object (Ricardo Martin)
  9. MSVC fixes (Petter Strandmark)
  10. Various fixes to (Markus Moll)
  11. Fixed a jacobian scaling bug.
  12. Numerically robust computation of model_cost_change.
  13. Signed comparison compiler warning fixes (Ricardo Martin)
  14. Various compiler warning fixes all over.
  15. Inclusion guard fixes (Petter Strandmark)
  16. Segfault in test code (Sergey Popov)
  17. Replaced EXPECT/ASSERT_DEATH with the more portable EXPECT_DEATH_IF_SUPPORTED macros.
  18. Fixed the camera projection model in Ceres’ implementation of Snavely’s camera model. (Ricardo Martin)


New Features

  1. Android Port (Scott Ettinger also contributed to the port)
  2. Windows port. (Changchang Wu and Pierre Moulon also contributed to the port)
  3. New subspace Dogleg Solver. (Markus Moll)
  4. Trust region algorithm now supports the option of non-monotonic steps.
  5. New loss functions ArcTanLossFunction, TolerantLossFunction and ComposedLossFunction. (James Roseborough).
  6. New DENSE_NORMAL_CHOLESKY linear solver, which uses Eigen’s LDLT factorization on the normal equations.
  7. Cached symbolic factorization when using CXSparse. (Petter Strandark)
  8. New example and data from the NIST non-linear regression test suite. (Thanks to Douglas Bates for suggesting this.)
  9. The traditional Dogleg solver now uses an elliptical trust region (Markus Moll)
  10. Support for returning initial and final gradients & Jacobians.
  11. Gradient computation support in the evaluators, with an eye towards developing first order/gradient based solvers.
  12. A better way to compute Solver::Summary::fixed_cost. (Markus Moll)
  13. CMake support for building documentation, separate examples, installing and uninstalling the library and Gerrit hooks (Arnaud Gelas)
  14. SuiteSparse4 support (Markus Moll)
  15. Support for building Ceres without TR1 (This leads to slightly slower DENSE_SCHUR and SPARSE_SCHUR solvers).
  16. BALProblem can now write a problem back to disk.
  17. bundle_adjuster now allows the user to normalize and perturb the problem before solving.
  18. Solver progress logging to file.
  19. Added Program::ToString and ParameterBlock::ToString to help with debugging.
  20. Ability to build Ceres as a shared library (MacOS and Linux only), associated versioning and build release script changes.
  21. Portable floating point classification API.

Bug Fixes

  1. Fix how invalid step evaluations are handled.
  2. Change the slop handling around zero for model cost changes to use relative tolerances rather than absolute tolerances.
  3. Fix an inadvertant integer to bool conversion. (Petter Strandmark)
  4. Do not link to libgomp when building on windows. (Petter Strandmark)
  5. Include gflags.h in (Petter Strandmark)
  6. Use standard random number generation routines. (Petter Strandmark)
  7. TrustRegionMinimizer does not implicitly negate the steps that it takes. (Markus Moll)
  8. Diagonal scaling allows for equal upper and lower bounds. (Markus Moll)
  9. TrustRegionStrategy does not misuse LinearSolver:Summary anymore.
  10. Fix Eigen3 Row/Column Major storage issue. (Lena Gieseke)
  11. QuaternionToAngleAxis now guarantees an angle in $[-pi, pi]$. (Guoxuan Zhang)
  12. Added a workaround for a compiler bug in the Android NDK to the Schur eliminator.
  13. The sparse linear algebra library is only logged in Summary::FullReport if it is used.
  15. Fix how static structure detection for the Schur eliminator logs its results.
  16. Correct example code in the documentation. (Petter Strandmark)
  17. Fix fpclassify.h to work with the Android NDK and STLport.
  18. Fix a memory leak in the
  19. Fix an early return bug in the Dogleg solver. (Markus Moll)
  20. Zero initialize Jets.
  21. Moved internal/ceres/mock_log.h to internal/ceres/gmock/mock-log.h
  22. Unified file path handling in tests.
  23. includes gflags
  24. Renamed Ceres’ Mutex class and associated macros to avoid namespace conflicts.
  25. Close the BAL problem file after reading it (Markus Moll)
  26. Fix IsInfinite on Jets.
  27. Drop alignment requirements for Jets.
  28. Fixed Jet to integer comparison. (Keith Leung)
  29. Fix use of uninitialized arrays. (Sebastian Koch & Markus Moll)
  30. Conditionally compile gflag dependencies.(Casey Goodlett)
  31. Add to the examples CMake file.


Bug Fixes

  1. suitesparse_test is enabled even when -DSUITESPARSE=OFF.
  2. FixedArray internal struct did not respect Eigen alignment requirements (Koichi Akabe & Stephan Kassemeyer).
  3. Fixed documentation and code mismatch (Nick Lewycky).


Bug Fixes

  1. Fix constant parameter blocks, and other minor fixes (Markus Moll)
  2. Fix alignment issues when combining Jet and FixedArray in automatic differeniation.
  3. Remove obsolete build_defs file.


New Features

  1. Powell’s Dogleg solver
  2. Documentation now has a brief overview of Trust Region methods and how the Levenberg-Marquardt and Dogleg methods work.

Bug Fixes

  1. Destructor for TrustRegionStrategy was not virtual (Markus Moll)
  2. Invalid DCHECK in (Markus Moll)
  3. Iteration callbacks were not properly invoked (Luis Alberto Zarrabeiti)
  4. Logging level changes in ConjugateGradientsSolver
  5. VisibilityBasedPreconditioner setup does not account for skipped camera pairs. This was debugging code.
  6. Enable SSE support on MacOS
  7. system_test was taking too long and too much memory (Koichi Akabe)


New Features

  1. CXSparse support.
  2. Block oriented fill reducing orderings. This reduces the factorization time for sparse CHOLMOD significantly.
  3. New Trust region loop with support for multiple trust region step strategies. Currently only Levenberg-Marquardt is supported, but this refactoring opens the door for Dog-leg, Stiehaug and others.
  4. CMake file restructuring. Builds in Release mode by default, and now has platform specific tuning flags.
  5. Re-organized documentation. No new content, but better organization.

Bug Fixes

  1. Fixed integer overflow bug in
  2. Renamed some macros to prevent name conflicts.
  3. Fixed incorrent input to StateUpdatingCallback.
  4. Fixes to AutoDiff tests.
  5. Various internal cleanups.


Bug Fixes

  1. Fix a bug in the handling of constant blocks. (Louis Simard)
  2. Add an optional lower bound to the Levenberg-Marquardt regularizer to prevent oscillating between well and ill posed linear problems.
  3. Some internal refactoring and test fixes.


New Features

  1. New iterative linear solver for general sparse problems - CGNR and a block Jacobi preconditioner for it.
  2. Changed the semantics of how SuiteSparse dependencies are checked and used. Now SuiteSparse is built by default, only if all of its dependencies are present.
  3. Automatic differentiation now supports dynamic number of residuals.
  4. Support for writing the linear least squares problems to disk in text format so that they can loaded into MATLAB.
  5. Linear solver results are now checked for nan and infinities.
  6. Added .gitignore file.
  7. A better more robust build system.

Bug Fixes

  1. Fixed a strict weak ordering bug in the schur ordering.
  2. Grammar and typos in the documents and code comments.
  3. Fixed tests which depended on exact equality between floating point values.


Initial open source release. Nathan Wiegand contributed to the Mac OSX port.


Ceres Solver grew out of the need for general least squares solving at Google. In early 2010, Sameer Agarwal and Fredrik Schaffalitzky started the development of Ceres Solver. Fredrik left Google shortly thereafter and Keir Mierle stepped in to take his place. After two years of on-and-off development, Ceres Solver was released as open source in May of 2012.