Implementation of the Levenberg-Marquardt optimization algorithm. More...

Inheritance diagram for OptimizationAlgorithmLevenberg< t_type >:

Collaboration diagram for OptimizationAlgorithmLevenberg< t_type >:

Public Member Functions
	OptimizationAlgorithmLevenberg ()
	construct the Levenberg algorithm, which will use the given Solver for solving the linearized system.
void	clear ()
	Resets all Levenberg-Marquardt state to initial values.
g_type	currentLambda () const
	return the currently used damping factor
int	levenbergIteration ()
	return the number of levenberg iterations performed in the last round
void	optimize (uint04 iterations, bool online, IndexScratch &scratch)
	Runs Levenberg-Marquardt optimization for the given number of iterations.
void	setUserLambdaInit (g_type lambda)
	Sets a user-specified initial damping factor.
OptimizationAlgorithm::SolverResult	solve (int iteration, bool online, IndexScratch &scratch)
	Performs a single Levenberg-Marquardt iteration with adaptive damping.
Public Member Functions inherited from OptimizationAlgorithmWithHessian< t_type >
	OptimizationAlgorithmWithHessian ()
	Default constructor.
void	addEdge (HyperGraph::HGEdge *e)
	Adds an edge to the optimizer.
void	addVertex (HyperGraph::HGVertex &v)
	Adds a vertex to the optimizer.
bool	buildLinearStructure (IndexScratch &scratch)
	Builds the linear system structure.
void	clear ()
	Clears the solver state.
bool	init (bool online=false)
	Initializes the solver, detecting whether Schur complement should be used.
SparseOptimizer &	optimizer ()
	Returns a mutable reference to the underlying sparse optimizer.
void	updateLinearSystem ()
	Rebuilds the linear system (Hessian and gradient) from the current graph state.
bool	updateStructure (const Buffer< HyperGraph::HGVertex * > &vset, const Set< HyperGraph::HGVertex * > &edges)
	Updates the linear system structure after graph topology changes.

Protected Member Functions
g_type	computeLambdaInit () const
	helper for Levenberg, this function computes the initial damping factor, if the user did not specify an own value, see setUserLambdaInit()
g_type	computeScale () const
	Computes the predicted reduction scale for the trust-region ratio.

Protected Attributes
g_type	_currentLambda
	Current damping factor.
g_type	_goodStepLowerScale
	Lower bound for lambda decrease if a good LM step.
g_type	_goodStepUpperScale
	Upper bound for lambda decrease if a good LM step.
g_type	_lambdaInit = 0.0
	User-specified initial lambda (0 = auto-compute).
int	_levenbergIterations
	The number of Levenberg iterations performed to accept the last step.
int	_nBad
	Counter for consecutive bad (non-improving) iterations.
g_type	_ni
	Multiplicative increase factor for lambda on bad steps.
g_type	_tau
	Scaling factor for initial lambda computation.

Additional Inherited Members
Public Types inherited from OptimizationAlgorithm
enum	SolverResult { Terminate =2 , OK =1 , Fail =-1 }
	Result codes returned by the solver after each iteration. More...
Public Attributes inherited from OptimizationAlgorithmWithHessian< t_type >
t_type	solver
	The underlying solver instance.

Detailed Description

template<class t_type>
class OptimizationAlgorithmLevenberg< t_type >

Implementation of the Levenberg-Marquardt optimization algorithm.

Template Parameters

t_type The block solver type.

Definition at line 10 of file optimization_algorithm_levenberg.h.

Member Function Documentation

◆ computeScale()

template<class t_type>

g_type OptimizationAlgorithmLevenberg< t_type >::computeScale ( ) const

inlineprotected

Computes the predicted reduction scale for the trust-region ratio.

Returns: The computed scale factor.

Definition at line 200 of file optimization_algorithm_levenberg.h.

References _currentLambda, and OptimizationAlgorithmWithHessian< t_solver >::solver.

Referenced by solve().

◆ optimize()

template<class t_type>

void OptimizationAlgorithmLevenberg< t_type >::optimize	(	uint04	iterations,
		bool	online,
		IndexScratch &	scratch )

inline

Runs Levenberg-Marquardt optimization for the given number of iterations.

Parameters

[in]	iterations	Maximum number of iterations.
[in]	online	Whether this is an online (incremental) optimization.
[in]	scratch	Scratch buffers for pose and landmark indices.

Definition at line 138 of file optimization_algorithm_levenberg.h.

References OptimizationAlgorithmWithHessian< t_solver >::init(), OptimizationAlgorithm::OK, OptimizationAlgorithmWithHessian< t_solver >::optimizer(), and solve().