Detailed Description

Class for graduation scheduling for riSAM.

Custom Schedulers can be implemented by defining custom Mu update strats. All loss functions are convex when $\mu = 0$ and recover their nonconvex form for $\mu = 1.0$.

For riSAM we recommend using GemanMcClure loss with SCALE_INVARIANT graduation scheme and the MuUpdateStable strategy configured by default for this class. Advanced users can explore alternate loss and graduation schedules.

Public Member Functions
Public Interface
	GraduationScheduler (MuUpdateStrategy muUpdateStrategy=MuUpdateStable, double muInitIncrement=0.2, double muInit=0.0)
	Individual Parameter Constructor.
double	muInit () const
	Returns the value of $\mu_{init}$ for this graduated.
double	updateMu (const double &mu, const double &residual, const size_t &updateCount) const
	Returns the next value of $\mu$ for this graduated.
double	updateMuInit (const double &muInit, const bool isInlier) const
	Returns the next value of $\mu_{init}$ for strong inliers/outliers.
bool	isMuConverged (const double &mu) const
	Returns true iff the value of $\mu$ has converged.

Static Public Member Functions
Default $\mu$ Update Strategies
static double	MuUpdateMcGann2023 (const double &mu, const double &residual, const size_t &updateCount)
	$\mu$ update sequence empirically discovered and presented in the orig riSAM paper
static double	MuUpdateStable (const double &mu, const double &residual, const size_t &updateCount)
	More stable $\mu$ Update sequence developed empirically since algorithm was published.

Public Types
Types
typedef std::shared_ptr< GraduationScheduler >	shared_ptr
	Shortcut for shared pointer.
typedef std::function< double(double, double, size_t)>	MuUpdateStrategy
	Function type for $\mu$ update sequence.

Protected Attributes
Fields
double	muInit_
	The initial value for mu $\mu_{init}$.
double	convergenceThreshold_
	The threshold at which to consider mu to be converged.
MuUpdateStrategy	muUpdateStrategy_
	The update strategy for the sequence $\mu$ values.
double	muInitIncrement_
	The amount to increment/decrement $\mu_{init}$ if the factor is a strong inlier/outlier when values converge.

Constructor & Destructor Documentation

◆ GraduationScheduler()

gtsam::GraduationScheduler::GraduationScheduler	(	MuUpdateStrategy	muUpdateStrategy = MuUpdateStable,
		double	muInitIncrement = 0.2,
		double	muInit = 0.0 )

inline

Individual Parameter Constructor.

Parameters

muUpdateStrategy	The update strategy to use for $\mu$ updates Recommend: MuUpdateStable, Alt: MuUpdateMcGann2023 Recs. to be used with GemanMcClure loss with SCALE_INVARIANT graduation
muInitIncrement	The amount to increment/decrement $\mu_{init}$
muInit	The starting value of $\mu$ for factors using this scheduler, in [0, 1]. 0 is the most convex start, and it also floors updateMuInit, so a nonzero value never graduates more convex than it.

Exceptions

std::invalid_argument if muInit is outside [0, 1].

Member Function Documentation

◆ MuUpdateMcGann2023()

double gtsam::GraduationScheduler::MuUpdateMcGann2023	(	const double &	mu,
		const double &	residual,
		const size_t &	updateCount )

static

$\mu$ update sequence empirically discovered and presented in the orig riSAM paper

Note: Adheres to MuUpdateStrategy interface

◆ MuUpdateStable()

double gtsam::GraduationScheduler::MuUpdateStable	(	const double &	mu,
		const double &	residual,
		const size_t &	updateCount )

static

More stable $\mu$ Update sequence developed empirically since algorithm was published.

Note: Adheres to MuUpdateStrategy interface

◆ updateMu()

double gtsam::GraduationScheduler::updateMu	(	const double &	mu,
		const double &	residual,
		const size_t &	updateCount ) const

Returns the next value of $\mu$ for this graduated.

Parameters

mu	The current value of $\mu$
residual	The current residual of the factor
updateCount	The number of mu updates during this solve

◆ updateMuInit()

double gtsam::GraduationScheduler::updateMuInit	(	const double &	muInit,
		const bool	isInlier ) const

Returns the next value of $\mu_{init}$ for strong inliers/outliers.

Inliers are updated to have more convex $\mu_{init}$
Outliers are update to have less convex $\mu_{init}$
Parameters

muInit The current value of $\mu_{init}$

isInlier Flag indicating inlier update otherwise an outlier update

Member Data Documentation

◆ convergenceThreshold_

double gtsam::GraduationScheduler::convergenceThreshold_

protected

The threshold at which to consider mu to be converged.

Note: $\mu = 1$ is the target robust loss by GTSAM convention

The documentation for this class was generated from the following files:

/tmp/gtsam-4.3.0-doxygen.rsXPUS/source/gtsam/sam/RISAMGraduationScheduler.h
/tmp/gtsam-4.3.0-doxygen.rsXPUS/source/gtsam/sam/RISAMGraduationScheduler.cpp

muInit	The current value of $\mu_{init}$
isInlier	Flag indicating inlier update otherwise an outlier update