Ceres

Ceres

The manif package differentiates Jacobians with respect to a local perturbation on the tangent space. These Jacobians map tangent spaces, as described in this paper.

However, many non-linear solvers (e.g. Ceres) expect functions to be differentiated with respect to the underlying representation vector of the group element (e.g. with respect to quaternion vector for SO3).

For this reason manif is compliant with Ceres auto-differentiation and the ceres::Jet type.

For reference of the size of the Jacobians returned when using ceres::Jet, manif implements rotations in the following way:

SO(2) and SE(2): as a complex number with real = cos(theta) and imag = sin(theta) values.
SO(3), SE(3) and SE_2(3): as a unit quaternion, using the underlying Eigen::Quaternion type.

Therefore, the respective Jacobian sizes using ceres::Jet are as follows:

ℝ(n) : size n
SO(2) : size 2
SO(3) : size 4
SE(2) : size 4
SE(3) : size 7
SE_2(3): size 10

Jacobians

Considering,

$\color{white}\bf x$

a group element (e.g. S3),

$\color{white}\omega$

the vector tangent to the group at

$\color{white}\bf x$

$\color{white}{\bf e}=f({\bf x})$

an error function,

one is interested in expressing the Taylor series of the error function,

$\color{white}f({\bf x\oplus\omega})\approx{\bf e}+{\bf J}_{\omega}^{e}~\omega .$

Therefore we have to compute the Jacobian

$\color{white}{\bf J}_{\omega}^{e}=\frac{\delta{\bf e}}{\delta{\bf x}}=\frac{\delta f({\bf x})}{\delta{\bf x}}=\lim_{\omega\to0}\frac{f({\bf x}\oplus\omega)\ominus f({\bf x})}{\omega}, (1)$

the Jacobian of

$\color{white}f({\bf x})$

with respect to a perturbation on the tangent space, so that the state update happens on the manifold tangent space.

In Ceres' framework, the computation of this Jacobian is decoupled in two folds as explained hereafter. The following terminology should sounds familiar to Ceres users.

Ceres CostFunction is the class representing and implementing a function

$\color{white}{\bf e}=f({\bf x})$

, for instance,

class QuadraticCostFunction : public ceres::SizedCostFunction<1, 1> {
 public:
  virtual ~QuadraticCostFunction() {}
  virtual bool Evaluate(double const* const* parameters,
                        double* residuals,
                        double** jacobians) const {
    const double x = parameters[0][0];
    residuals[0] = 10 - x;

    // Compute the Jacobian if asked for.
    if (jacobians != NULL && jacobians[0] != NULL) {
      jacobians[0][0] = -1;
    }
    return true;
  }
};

It produces the Jacobian,

$\color{white}{\bf J}_{{\bf x}\oplus\omega}^{e}=\frac{\delta{\bf e}}{\delta({\bf x}\oplus\omega)}=\lim_{\mathbf h\to0}\frac{ f({\bf x}+\mathbf h)-f({\bf x})}{\mathbf h}. (2)$

In Ceres, a LocalParameterization can be associated to a state.

Eigen::Quaterniond my_state;

ceres::Problem::Options problem_options;
ceres::Problem problem(problem_options);

// Add the state to Ceres problem
problem->AddParameterBlock(my_state.data(), 4);

// Associate a LocalParameterization to the state vector
problem_->SetManifold(my_state.data(),
                      new EigenQuaternionParameterization() );

The LocalParameterization class (and derived) performs the state update step of the optimization - the

$\color{white}{\bf x}\oplus\mathbf\omega$

operation. While the function operates for any

$\color{white}\mathbf\omega$

, its Jacobian is evaluated for

$\color{white}\omega=0$

thus providing the Jacobian,

$\color{white}{\bf J}_{\omega}^{{\bf x}\oplus\omega}=\frac{\delta({\bf x}\oplus\omega)}{\delta\omega}=\lim_{\delta\omega\to0}\frac{{\bf x}\oplus(\omega+\delta\omega)-{\bf x}\oplus\mathbf\omega}{\delta\omega}=\lim_{\delta\omega\to0}\frac{{\bf x}\oplus\delta\omega-{\bf x}}{\delta\omega}. (3)$

Once both the CostFunction and LocalParameterization's Jacobians are evaluated, Ceres internally computes (1) with the following product,

$\color{white}{\bf J}_{\omega}^{e}={\bf J}_{{\bf x}\oplus\omega}^{e}\times{\bf J}_{\omega}^{{\bf x}\oplus\omega}. (4)$

Voila.

The intermediate Jacobians (2-3) that Ceres requires are not available in manif since it provide directly the final Jacobian detailed in (1).

However, one still wants to use manif with his Ceres-based project. For this reason, manif is compliant with Ceres auto-differentiation and the ceres::Jet type to compute (2-3).

Below are presented two small examples illustrating how manif can be used with Ceres.

Example : A group-abstract LocalParameterization

Is shown here how one can implement a ceres::LocalParameterization-derived class using manif, that does the

$\color{white}{\bf x}\oplus\mathbf\omega$

for any group implemented in manif passed as a template parameter.

template <typename _LieGroup>
class CeresLocalParameterization
{
  using LieGroup = _LieGroup;
  using Tangent  = typename _LieGroup::Tangent;

  template <typename _Scalar>
  using LieGroupTemplate = typename LieGroup::template LieGroupTemplate<_Scalar>;

  template <typename _Scalar>
  using TangentTemplate = typename Tangent::template TangentTemplate<_Scalar>;

public:

  CeresLocalParameterizationFunctor() = default;
  virtual ~CeresLocalParameterizationFunctor() = default;

  template<typename T>
  bool operator()(const T* state_raw,
                  const T* delta_raw,
                  T* state_plus_delta_raw) const
  {
    const Eigen::Map<const LieGroupTemplate<T>> state(state_raw);
    const Eigen::Map<const TangentTemplate<T>>  delta(delta_raw);

    Eigen::Map<LieGroupTemplate<T>> state_plus_delta(state_plus_delta_raw);

    state_plus_delta = state + delta;

    return true;
  }
};
//
...
// Some typedef helpers
using CeresLocalParameterizationSO2 = CeresLocalParameterizationFunctor<SO2d>;
using CeresLocalParameterizationSE2 = CeresLocalParameterizationFunctor<SE2d>;
using CeresLocalParameterizationSO3 = CeresLocalParameterizationFunctor<SO3d>;
using CeresLocalParameterizationSE3 = CeresLocalParameterizationFunctor<SE3d>;

Example : A small Ceres problem

This example highlights the use of the predefined Ceres helper classes available with manif. In this example, we compute an average point from 4 points in SE2.

// Tell ceres not to take ownership of the raw pointers
ceres::Problem::Options problem_options;
problem_options.cost_function_ownership = ceres::DO_NOT_TAKE_OWNERSHIP;
problem_options.manifold_ownership = ceres::DO_NOT_TAKE_OWNERSHIP;

ceres::Problem problem(problem_options);

// We use a first manif helper that creates a ceres cost-function.
// The cost function computes the distance between
// the desired state and the current state

// Create 4 objectives which are 'close' in SE2.
std::shared_ptr<ceres::CostFunction> obj_pi_over_4   = manif::make_objective_autodiff<SE2d>(3, 3,    M_PI/4.);
std::shared_ptr<ceres::CostFunction> obj_3_pi_over_8 = manif::make_objective_autodiff<SE2d>(3, 1, 3.*M_PI/8.);
std::shared_ptr<ceres::CostFunction> obj_5_pi_over_8 = manif::make_objective_autodiff<SE2d>(1, 1, 5.*M_PI/8.);
std::shared_ptr<ceres::CostFunction> obj_3_pi_over_4 = manif::make_objective_autodiff<SE2d>(1, 3, 3.*M_PI/4.);

SE2d average_state(0,0,0);


// Add residual blocks to ceres problem
problem.AddResidualBlock( obj_pi_over_4.get(),
                          nullptr,
                          average_state.data() );

problem.AddResidualBlock( obj_3_pi_over_8.get(),
                          nullptr,
                          average_state.data() );

problem.AddResidualBlock( obj_5_pi_over_8.get(),
                          nullptr,
                           average_state.data() );

problem.AddResidualBlock( obj_3_pi_over_4.get(),
                          nullptr,
                          average_state.data() );

// We use a second manif helper that creates a ceres local parameterization
// for our optimized state block.

std::shared_ptr<ceres::Manifold>
  auto_diff_manifold =
    manif::make_manifold_autodiff<SE2d>();

problem.SetManifold( average_state.data(),
                     auto_diff_manifold.get() );

// Run the solver!
ceres::Solver::Options options;
options.minimizer_progress_to_stdout = true;

ceres::Solver::Summary summary;
ceres::Solve(options, &problem, &summary);

std::cout << "summary:\n" << summary.FullReport() << "\n";

std::cout << "Average state:\nx:" << average_state.x()
          << "\ny:" << average_state.y()
          << "\nt:" << average_state.angle()
          << "\n\n";