Train and use an SVM to predict antipodal grasps. More...

#include <learning.h>

Classes
struct	Instance
	Learning instance representing a grasp hypothesis. More...
struct	UniqueVectorComparator
	Comparator for 2D vectors. More...
Public Member Functions
std::vector< GraspHypothesis >	classify (const std::vector< GraspHypothesis > &hands_list, const std::string &svm_filename, const Eigen::Matrix3Xd &cam_pos, bool is_plotting=false)
	Predict antipodal grasps using an SVM classifier.
	Learning ()
	Constructor.
	Learning (int num_threads)
	Constructor. Set the number of threads to be used for prediction.
void	train (const std::vector< GraspHypothesis > &hands_list, const std::vector< int > &sizes, const std::string &file_name, const Eigen::Matrix3Xd &cam_pos, int max_positive=1000000000, bool is_plotting=false)
	Train an SVM classifier to predict antipodal grasps.
void	train (const std::vector< GraspHypothesis > &hands_list, const std::string &file_name, const Eigen::Matrix3Xd &cam_pos, bool is_plotting=false)
	Train an SVM classifier to predict antipodal grasps.
void	trainBalanced (const std::vector< GraspHypothesis > &hands_list, const std::vector< int > &sizes, const std::string &file_name, const Eigen::Matrix3Xd &cam_pos, int max_positive=1000000000, bool is_plotting=false)
	Train an SVM classifier to predict antipodal grasps. Uses the same number of positives and negatives for each training point cloud.
Private Member Functions
void	convertData (const std::vector< Instance > &instances, const std::string &file_name, bool is_plotting=false, bool uses_linear_kernel=false)
	Convert a set of learning instances to training data readable by the SVM.
cv::Mat	convertToImage (const Instance &ins)
	Convert a learning instance to a grasp image.
Instance	createInstance (const GraspHypothesis &h, const Eigen::Matrix3Xd &cam_pos, int cam=-1)
	Create a learning instance from a grasp hypothesis.
Eigen::VectorXi	floorVector (const Eigen::VectorXd &a)
	Round a vector's elements down to the closest, smaller integers.
Private Attributes
int	num_horizontal_cells_
	the width of a grasp image
int	num_threads_
	the number of threads used for prediction
int	num_vertical_cells_
	the height of a grasp image

Detailed Description

Train and use an SVM to predict antipodal grasps.

Learning class

This class trains an SVM classifier to predict antipodal grasps. Once trained, it can be used to predict antipodal grasps. The classifier is trained with HOG features obtained from grasp images. A grasp image is a 2D image representation of a grasp hypothesis.

Definition at line 56 of file learning.h.

Constructor & Destructor Documentation

Learning::Learning ( ) [inline]

Constructor.

Definition at line 63 of file learning.h.

Learning::Learning ( int num_threads ) [inline]

Constructor. Set the number of threads to be used for prediction.

Definition at line 71 of file learning.h.

Member Function Documentation

std::vector< GraspHypothesis > Learning::classify	(	const std::vector< GraspHypothesis > &	hands_list,
		const std::string &	svm_filename,
		const Eigen::Matrix3Xd &	cam_pos,
		bool	is_plotting = `false`
	)

Predict antipodal grasps using an SVM classifier.

Parameters:

hands_list	the set of grasp hypotheses to be tested for being antipodal grasps
file_name	the location and name of the file in which the SVM is stored
cam_pos	the camera poses
is_plotting	whether the training process is visualized

Definition at line 165 of file learning.cpp.

void Learning::convertData	(	const std::vector< Instance > &	instances,
		const std::string &	file_name,
		bool	is_plotting = `false`,
		bool	uses_linear_kernel = `false`
	)		`[private]`

Convert a set of learning instances to training data readable by the SVM.

Parameters:

instances	the training instances to be converted
file_name	the location and name of the file in which the SVM is stored
is_plotting	whether the conversion process is visualized
uses_linear_kernel	whether a linear or a quadratic kernel is used in the SVM

Definition at line 249 of file learning.cpp.

cv::Mat Learning::convertToImage ( const Instance & ins ) [private]

Convert a learning instance to a grasp image.

Parameters:

ins	the learning instance to be converted

Returns:: the created image

Definition at line 320 of file learning.cpp.

Learning::Instance Learning::createInstance	(	const GraspHypothesis &	h,
		const Eigen::Matrix3Xd &	cam_pos,
		int	cam = `-1`
	)		`[private]`

Create a learning instance from a grasp hypothesis.

Parameters:

h	the grasp hypothesis from which the learning instance is created
cam_pos	the camera poses
cam	index of the camera which produced the points in the grasp hypothesis

Returns:: the created learning instance

Definition at line 375 of file learning.cpp.

Eigen::VectorXi Learning::floorVector ( const Eigen::VectorXd & a ) [private]

Round a vector's elements down to the closest, smaller integers.

Parameters:

a	the vector whose elements are rounded down

Returns:: the vector containing the rounded elements

Definition at line 367 of file learning.cpp.

void Learning::train	(	const std::vector< GraspHypothesis > &	hands_list,
		const std::vector< int > &	sizes,
		const std::string &	file_name,
		const Eigen::Matrix3Xd &	cam_pos,
		int	max_positive = `1000000000`,
		bool	is_plotting = `false`
	)

Train an SVM classifier to predict antipodal grasps.

Parameters:

hands_list	the set of grasp hypotheses to be used for training
sizes	the number of grasp hypotheses found for each training point cloud
file_name	the location and name of the file in which the SVM is stored
max_positive	the maximum number of positives examples to be used from each training point cloud
cam_pos	the camera poses
is_plotting	whether the training process is visualized

Definition at line 76 of file learning.cpp.

void Learning::train	(	const std::vector< GraspHypothesis > &	hands_list,
		const std::string &	file_name,
		const Eigen::Matrix3Xd &	cam_pos,
		bool	is_plotting = `false`
	)

Train an SVM classifier to predict antipodal grasps.

Parameters:

hands_list	the set of grasp hypotheses to be used for training
sizes	the number of grasp hypotheses found for each training point cloud
file_name	the location and name of the file in which the SVM is stored
is_plotting	whether the training process is visualized

Definition at line 143 of file learning.cpp.

void Learning::trainBalanced	(	const std::vector< GraspHypothesis > &	hands_list,
		const std::vector< int > &	sizes,
		const std::string &	file_name,
		const Eigen::Matrix3Xd &	cam_pos,
		int	max_positive = `1000000000`,
		bool	is_plotting = `false`
	)

Train an SVM classifier to predict antipodal grasps. Uses the same number of positives and negatives for each training point cloud.

Parameters:

hands_list	the set of grasp hypotheses to be used for training
sizes	the number of grasp hypotheses found for each training point cloud
file_name	the location and name of the file in which the SVM is stored
cam_pos	the camera poses
max_positive	the maximum number of positives examples to be used from each training point cloud
is_plotting	whether the training process is visualized