#include <dataset.h>

Inheritance diagram for ebl::dataset< Tdata >:

Public Member Functions
	dataset (const char name, const char inroot=NULL)
virtual	~dataset ()
	Destructor.
bool	alloc (intg max=0)
virtual bool	extract ()
	Extract data from files into dataset.
virtual void	extract_statistics ()
	Extracts statistics about dataset into outdir/stats.csv.
bool	split_max_and_save (const char name1, const char name2, intg max, const string &outroot)
void	split_max (dataset< Tdata > &ds1, dataset< Tdata > &ds2, intg max)
void	merge_and_save (const char name1, const char name2, const string &outroot)
void	shuffle ()
	Shuffle order of samples.
virtual void	set_unique_label (const string &class_name)
	Set a unique label 'class_name' to all samples.
void	set_planar_loading ()
	Input is loaded as planar (channels in first dimension).
const idxdim &	get_sample_outdim ()
	Get sample dimensions.
intg	size ()
	Return the number of samples in the dataset.
t_label	get_label_from_class (const string &class_name)
	Return unsigned int label corresponding to class name.
void	set_display (bool display)
	Turn display on or off.
void	set_sleepdisplay (uint delay)
	Make the display sleep delay milliseconds between frames.
void	set_preprocessing (vector< resizepp_module< fs(Tdata)> * > &p)
	Set a vector for preprocessing modules.
virtual void	set_outdims (const idxdim &d)
virtual void	set_outdir (const char s, const char tmp=NULL)
	Specify output directory 's' and a temporary output directory 'tmp'.
void	set_mindims (const idxdim &d)
void	set_maxdims (const idxdim &d)
	Specify the maximum dimensions of input samples.
void	set_scales (const vector< double > &sc, const string &od)
void	set_fovea (const vector< double > &scales)
void	set_max_per_class (intg max)
	Set all max per class to max.
void	set_max_data (intg max)
	Set maximum number of samples.
void	set_image_pattern (const string &p)
	Set the pattern used to find image.
void	set_exclude (const vector< string > &ex)
	Set the list of classes to exclude (including everything else).
void	set_include (const vector< string > &inc)
	Set the list of classes to include (excluding everything else).
void	set_save (const string &save)
void	set_individual_save (bool b)
	If 'b' is true, save each intermediate sample file individually.
void	set_separate_layers_save (bool b)
	If 'b' is true, save each layer of each sample individually.
void	set_name (const string &name)
void	set_label (const string &label)
void	set_bbox_woverh (float factor)
void	set_nopadded (bool nopadded)
void	set_videobox (uint nframes, uint stride)
void	set_jitter (uint tjitter_step, uint tjitter_hmin, uint tjitter_hmax, uint tjitter_wmin, uint tjitter_wmax, uint scale_steps, float scale_min, float scale_max, uint rotation_steps, float rotation_range, uint njitter)
virtual void	set_minvisibility (float minvis)
void	set_wmirror ()
	Add sample mirrored with vertical-axis symmetry.
void	save_display (const string &dir, uint h=0, uint w=0)
void	use_pose ()
void	use_parts ()
void	use_parts_only ()
bool	full (t_label label=-1)
virtual intg	count_total ()
bool	load (const string &root)
bool	save (const string &root, bool save_data=true)
void	print_classes ()
	Prints number of classes and list on standard output.
void	print_stats ()
	Prints number of samples for each class on std output.
Static Public Member Functions
static idx< ubyte >	build_classes_idx (vector< string > &classes)
	Returns a matrix of class names based on the classes vector.
Protected Member Functions
bool	allocate (intg n, idxdim &d)
virtual bool	add_data (midx< Tdata > &d, const t_label label, const string class_name, const char filename=NULL, const rect< int > r=NULL, pair< int, int > center=NULL, const rect< int > visr=NULL, const rect< int > cropr=NULL, const vector< object * > objs=NULL, const jitter jittforce=NULL)
void	add_data2 (midx< Tdata > &sample, t_label label, const string class_name, const char filename, const jitter jitt, idx< t_jitter > js)
	add/save sample, called at the end of add_data().
void	add_label (t_label label, const string class_name, const char filename, const jitter jitt, idx< t_jitter > js)
	Add/save label, called by add_data2().
virtual void	clear_classes ()
	Clear the classes.
virtual bool	add_class (const string &class_name)
	add a class name
virtual void	set_classes (idx< ubyte > &classidx)
	set all classes directly using an idx of classes strings
virtual intg	count_samples ()
	Count and returns how many samples are present in data.
void	split (dataset< Tdata > &ds1, dataset< Tdata > &ds2)
template<class Toriginal >
bool	save_scales (idx< Toriginal > &d, const string &filename)
virtual bool	included (t_label &lab)
virtual bool	included (const string &class_name)
midx< Tdata >	preprocess_data (midx< Tdata > &d, const string class_name, const char filename=NULL, const rect< int > r=NULL, double scale=0, rect< int > outr=NULL, pair< int, int > center=NULL, jitter jitt=NULL, const rect< int > visr=NULL, const rect< int > cropr=NULL, rect< int > *inr_out=NULL)
void	display_added (midx< Tdata > &added, idx< Tdata > &original, const string class_name, const char filename=NULL, const rect< int > inr=NULL, const rect< int > origr=NULL, bool active_sleepd=true, pair< int, int > center=NULL, const rect< int > visr=NULL, const rect< int > cropr=NULL, const vector< object > objs=NULL, const jitter jitt=NULL, idx< t_jitter > js=NULL, uint woriginal=NULL)
string &	get_class_string (t_label id)
	Return the class name associated with label id.
t_label	get_class_id (const string &name)
	Return the label id associated with class name.
void	compute_stats ()
	Compute statistics about the dataset.
uint	count_matches (const string &dir, const string &pattern)
	Count (recursively) how many files matching pattern are inside dir.
void	process_dir (const string &dir, const string &ext, const string &class_name)
virtual void	load_data (const string &fname)
	Method to load an image.
virtual void	compute_random_jitter ()
Protected Attributes
midx< Tdata >	data
	data matrix
idx< t_label >	labels
	labels matrix
idx< intg >	ids
	indices matrix
midx< t_jitter >	jitters
	jitter info matrix
vector< string >	classes
	list of classes strings
idx< t_label >	classpairs
	sample pairs class-wise
idx< t_label >	deformpairs
	sample pairs deformation-wise
uint	height
	height of output
uint	width
	width of output
bool	allocated
	data matrices allocated or not
bool	no_outdims
	no outdims were specified.
idxdim	outdims
	dims of sample out dimensions
idxdim	mindims
	min dims of input samples
idxdim	maxdims
	max dims of input samples
bool	maxdims_set
	max dims of input samples is set?
idxdim	datadims
	dimensions of data out dimensions
uint	nlayers
	# layers per sample
intg	data_cnt
	number of samples added so far
intg	processed_cnt
	#processed (not necessarly added)
intg	max_data
	user can limit samples# with this
bool	max_data_set
	max_data been set by user or not
intg	total_samples
	number of samples of dataset
idx< intg >	max_per_class
	max # samples per class
intg	mpc
	value to put in max_per_class
bool	max_per_class_set
	mpc has been set or not
midx< Tdata >	load_img
	temporary image loader
bool	scale_mode
	scales saving mode
vector< double >	scales
	integer scales
bool	interleaved_input
	indicate input is interleaved
bool	load_planar
	input is planar when loaded
vector< string >	exclude
	list of excluded classes
vector< string >	include
	list of included classes
bool	usepose
	use pose or not
bool	useparts
	use parts or not
bool	usepartsonly
	use parts only or not
string	save_mode
	saving mode (dataset, ppm, png..)
bool	individual_save
	Save individual files or not.
bool	separate_layers_save
	Save separate layers or not.
list< string >	images_list
	List of saved image files.
bool	wmirror
	add vertical-axis symmetry
float	bbox_woverh
	bounding boxes width over h factor
string	force_label
	force all labels to this one
bool	nopadded
	ignore too small samples
float	minvisibility
int	tjitter_step
	Translation step, in pixels.
int	tjitter_hmin
	Translation height jitter min.
int	tjitter_hmax
	Translation height jitter max.
int	tjitter_wmin
	Translation width jitter min.
int	tjitter_wmax
	Translation width jitter max.
int	sjitter_steps
	number of possible scales
float	sjitter_min
	min range of scales
float	sjitter_max
	max range of scales
int	rjitter_steps
	number of possible rotations
float	rjitter
	range of rotations
uint	njitter
	number of random jitters
bool	bjitter
	use jitter or not
vector< jitter >	random_jitter
	pre-computed list of random jitter
string	name
	dataset name
string	data_fname
	data filename
string	labels_fname
	labels filename
string	jitters_fname
	jitters filename
string	ids_fname
	ids filename
string	classes_fname
	classes filename
string	classpairs_fname
	classpairs filename
string	deformpairs_fname
	deformpairs filename
string	inroot
	root directory of input files
string	outdir
	root directory of output files
string	outtmp
	root directory of tmp output files
string	extension
	extension of files to extract
bool	display_extraction
	display during extraction
bool	display_result
	display extracted dataset
bool	sleep_display
	enable sleeping when displaying
uint	sleep_delay
	display sleep delay in ms
bool	bsave_display
string	save_display_dir
uint	nclasses
	Number of classes.
idx< intg >	class_tally
	counter for class tally
idx< intg >	add_tally
	counter for additions tally
uint	add_errors
	Number of adding failures.
timer	xtimer
	Extraction timer.
bool	do_preprocessing
	activate or deactivate pp
vector< resizepp_module< fs(Tdata)> * >	ppmods
	pp resizing module
string	pp_names
	All preprocessing names.
rect< int >	original_bbox
	bbox of image in resized image
vector< double >	fovea
	fovea context
bool	do_videobox
	Activate or deactivate videobox.
uint	videobox_nframes
	Number of future frames.
uint	videobox_stride
	Stride of iterating over frames.

Detailed Description

template<class Tdata>
class ebl::dataset< Tdata >

The dataset class allows to extract a dataset from sample files and compile all samples into one dataset matrix, formatted for learning.

Constructor & Destructor Documentation

template<class Tdata >

ebl::dataset< Tdata >::dataset	(	const char *	name,
		const char *	inroot = `NULL`
	)

Initialize the dataset's name and other internal variables, but does not allocate data matrices, user must call alloc for that effect. outdims are the target output dimensions of each sample. inroot is the root directory where to extract data.

Member Function Documentation

template<class Tdata >

bool ebl::dataset< Tdata >::add_data	(	midx< Tdata > &	d,
		const t_label	label,
		const string *	class_name,
		const char *	filename = `NULL`,
		const rect< int > *	r = `NULL`,
		pair< int, int > *	center = `NULL`,
		const rect< int > *	visr = `NULL`,
		const rect< int > *	cropr = `NULL`,
		const vector< object * > *	objs = `NULL`,
		const jitter *	jittforce = `NULL`
	)		`[protected, virtual]`

add sample d to the data with label class_name (and converting from Toriginal to Tdata type). r is an optional region of interest rectangle in the image d.

Parameters:

visr	An optional bounding box for the visible area of the object
cropr	An optional bounding box of ROI of the image to crop. Image should already be cropped, use only for display.
objs	An optional vector of other objects in current image, for display purposes.
objs	An optional vector of other objects in current image, for display purposes.
jittforce	If not null, ignore other jittering and use only this one.

template<class Tdata >

bool ebl::dataset< Tdata >::alloc ( intg max = 0 )

Count how many samples are present in files to be compiled, and allocate buffers to the number of samples. If max is given (default is 0), then the dataset is limited to max samples (or counted samples if less).

template<class Tdata >

bool ebl::dataset< Tdata >::allocate	(	intg	n,
		idxdim &	d
	)		`[protected]`

Allocate dataset with n samples, each with dimensions d. This method is not called by the constructor and must be called before adding data samples. The buffers can be allocated only once, this is not meant to be used as a buffer resize method.

template<class Tdata >

void ebl::dataset< Tdata >::compute_random_jitter ( ) [protected, virtual]

Fills internal 'random_jitter' vector with all possible jitters in random order.

Reimplemented in ebl::pascal_dataset< Tdata >.

template<class Tdata >

intg ebl::dataset< Tdata >::count_total ( ) [virtual]

Count how many samples will be present in dataset files to be compiled, this may be more than count_samples() when jitter or mirroring is activated. This returns the total.

template<class Tdata >

void ebl::dataset< Tdata >::display_added	(	midx< Tdata > &	added,
		idx< Tdata > &	original,
		const string *	class_name,
		const char *	filename = `NULL`,
		const rect< int > *	inr = `NULL`,
		const rect< int > *	origr = `NULL`,
		bool	active_sleepd = `true`,
		pair< int, int > *	center = `NULL`,
		const rect< int > *	visr = `NULL`,
		const rect< int > *	cropr = `NULL`,
		const vector< object * > *	objs = `NULL`,
		const jitter *	jitt = `NULL`,
		idx< t_jitter > *	js = `NULL`,
		uint *	woriginal = `NULL`
	)		`[protected]`

Display the added sample and the original image.

Parameters:

visr	An optional bounding box for the visible area of the object
cropr	An optional bounding box of ROI of the image to crop. Image should already be cropped, use only for display.
scale	Scale the input box by this factor.
inr	The actual rectangle used for extraction.
origr	The original object bounding box.
woriginal	If defined, it gets filled with the width offset at which the original image gets painted.

template<class Tdata >

bool ebl::dataset< Tdata >::full ( t_label label = -1 )

Dataset has reached maximum sample capacity (this can be controlled by setting max_data variable). A label can be optionally passed to return if the dataset is full for a particular class (this is relevant only if set_max_per_class has been previously set.

template<class Tdata >

bool ebl::dataset< Tdata >::included ( t_label & lab ) [protected, virtual]

Returns true if label is authorized (based on excluded and included variables).

template<class Tdata >

bool ebl::dataset< Tdata >::included ( const string & class_name ) [protected, virtual]

Returns true if class_name is authorized (based on excluded and included variables).

template<class Tdata >

bool ebl::dataset< Tdata >::load ( const string & root )

Load dataset found in root. root should be including the name of the dataset, e.g.: "/data/dataset1" is going to load dataset1_*.mat

template<class Tdata >

void ebl::dataset< Tdata >::merge_and_save	(	const char *	name1,
		const char *	name2,
		const string &	outroot
	)

merge datasets with names name1 and name2 found in outroot into this dataset and save it.

template<class Tdata >

midx< Tdata > ebl::dataset< Tdata >::preprocess_data	(	midx< Tdata > &	d,
		const string *	class_name,
		const char *	filename = `NULL`,
		const rect< int > *	r = `NULL`,
		double	scale = `0`,
		rect< int > *	outr = `NULL`,
		pair< int, int > *	center = `NULL`,
		jitter *	jitt = `NULL`,
		const rect< int > *	visr = `NULL`,
		const rect< int > *	cropr = `NULL`,
		rect< int > *	inr_out = `NULL`
	)		`[protected]`

Preprocess data d of type Toriginal into an idx of type Tdata with output dimensions outdims and return the result. The type of preprocessing can be selected using set_pp_conversion().

Parameters:

outr	If not null, copy the rect of the input region in the output image.
visr	An optional bounding box for the visible area of the object
cropr	An optional bounding box of ROI of the image to crop. Image should already be cropped, use only for display.
scale	Scale the input box by this factor.
inr_out	If not null, this rectangle will be filled with the actual rectangle used for extraction.

template<class Tdata >

void ebl::dataset< Tdata >::process_dir	(	const string &	dir,
		const string &	ext,
		const string &	class_name
	)		`[protected]`

Recursively add all files matching the pattern ext to the dataset with class class_name.

template<class Tdata >

bool ebl::dataset< Tdata >::save	(	const string &	root,
		bool	save_data = `true`
	)

Save dataset in root.

Parameters:

save_data If false, do not save data matrix.

template<class Tdata >

void ebl::dataset< Tdata >::save_display	(	const string &	dir,
		uint	h = `0`,
		uint	w = `0`
	)

Saves all displayed frames to 'dir'. If h and w are different than 0, fix the saved frames to hxw.

template<class Tdata >

void ebl::dataset< Tdata >::set_bbox_woverh ( float factor )

Force bounding box width to be h * factor. This is useful to normalize bounding boxes with varying width to a fixed height/width ratio.

template<class Tdata >

void ebl::dataset< Tdata >::set_fovea ( const vector< double > & scales )

Setting fovea scales: the feature layer is duplicated for each scale factor provided here.

template<class Tdata >

void ebl::dataset< Tdata >::set_jitter	(	uint	tjitter_step,
		uint	tjitter_hmin,
		uint	tjitter_hmax,
		uint	tjitter_wmin,
		uint	tjitter_wmax,
		uint	scale_steps,
		float	scale_min,
		float	scale_max,
		uint	rotation_steps,
		float	rotation_range,
		uint	njitter
	)

Add n samples randomly jittered over in a (minradius,maxradius) spatial neighborhood around original location.

template<class Tdata >

void ebl::dataset< Tdata >::set_label ( const string & label )

If called, this method will force the dataset to assign this one label to all collected images.

template<class Tdata >

void ebl::dataset< Tdata >::set_mindims ( const idxdim & d )

Specify the minimum dimensions of input samples. The default is: 1x1

template<class Tdata >

void ebl::dataset< Tdata >::set_minvisibility ( float minvis ) [virtual]

Set minimum visibility ratio (between 0.0 and 1.0) of a sample. This may be used to ignore sample rects that are cropped too much while jittering. It may also be used when visible regions are defined, in which case the ratio is computed using the overlap of the visible bounding box with the original bounding box.

template<class Tdata >

void ebl::dataset< Tdata >::set_name ( const string & name )

Set name of dataset. This name will be used by load and save methods. It is usually set in the constructor, but one can also use this method instead.

template<class Tdata >

void ebl::dataset< Tdata >::set_nopadded ( bool nopadded )

If true, ignore samples with padded areas, i.e. too small for target size.

template<class Tdata >

void ebl::dataset< Tdata >::set_outdims ( const idxdim & d ) [virtual]

Specify the dimensions of 1 output sample. The default is: 96x96x3

template<class Tdata >

void ebl::dataset< Tdata >::set_save ( const string & save )

Set saving mode, default is "dataset", saving everything in single lush matrix format. other values can be "ppm", "jpg", etc.

template<class Tdata >

void ebl::dataset< Tdata >::set_scales	(	const vector< double > &	sc,
		const string &	od
	)

Setting scale mode and scales: preprocess and save each image in each scale in outdir directory.

template<class Tdata >

void ebl::dataset< Tdata >::set_videobox	(	uint	nframes,
		uint	stride
	)

Extract temporal features in video sequences by extracting the bboxes over n multiple frames(with a stride)

template<class Tdata >

void ebl::dataset< Tdata >::split	(	dataset< Tdata > &	ds1,
		dataset< Tdata > &	ds2
	)		`[protected]`

split current dataset into ds1 and ds2, using their max_per_class array (first filling ds1 until full, then ds2, therefore assuming that maximums are lower for ds1 than ds2).

template<class Tdata >

void ebl::dataset< Tdata >::split_max	(	dataset< Tdata > &	ds1,
		dataset< Tdata > &	ds2,
		intg	max
	)

Split dataset into datasets ds1 and ds2, by limiting ds1 to max samples per class, putting anything left into ds2.

template<class Tdata >

bool ebl::dataset< Tdata >::split_max_and_save	(	const char *	name1,
		const char *	name2,
		intg	max,
		const string &	outroot
	)

Split dataset into datasets ds1 and ds2, by limiting ds1 to max samples per class, putting anything left into ds2. Then save each dataset in outroot directory.

template<class Tdata >

void ebl::dataset< Tdata >::use_parts ( )

also extract parts of objects if available. e.g. in pascal if for class "person", parts "face", "hand" and "foot" are available.

template<class Tdata >

void ebl::dataset< Tdata >::use_parts_only ( )

only extract parts of objects if available. e.g. in pascal if for class "person", parts "face", "hand" and "foot" are available.

template<class Tdata >

void ebl::dataset< Tdata >::use_pose ( )

use pose information to separate classes. e.g. if for class "person" we have "front" and "side" pose, create 2 classes "person_front" and "person_side" instead of 1 class "person".

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

/home/rex/ebltrunk/tools/libeblearntools/include/dataset.h
/home/rex/ebltrunk/tools/libeblearntools/include/dataset.hpp

Public Member Functions

Static Public Member Functions

Protected Member Functions

Protected Attributes

Detailed Description

template<class Tdata> class ebl::dataset< Tdata >

Constructor & Destructor Documentation

Member Function Documentation

template<class Tdata>
class ebl::dataset< Tdata >