ebl::detector< T, Tstate > Class Template Reference

List of all members.

Public Member Functions
	detector (module_1_1< T, Tstate > &thenet, vector< string > &labels, answer_module< T, T, T, Tstate > *answer=NULL, resizepp_module< T, Tstate > *resize=NULL, const char *background=NULL, std::ostream &out=std::cout, std::ostream &err=std::cerr, bool adapt_scales=false)
virtual	~detector ()
	Destructor.
void	set_scaling_original ()
	Set the multi-scale to 1 scale only: the image's original scale.
void	set_scaling_type (t_scaling type)
	Set the scaling type.
void	set_resolutions (const midxdim &scales)
	Set all scales manually.
void	set_resolutions (const vector< double > &factors)
	Set scales a factors of the input sizes.
void	set_resolution (double factor)
	Set 1 scale only as a factor of the input sizes.
void	set_resolutions (int resolutions)
void	set_resolutions (idx< uint > &resolutions)
void	set_zpads (float hzpad, float wzpad)
void	set_resolutions (double scales_steps, double max_scale=1.0, double min_scale=1.0)
int	get_class_id (const string &name)
	Return the id of the class 'name' or -1 if not found.
void	set_bgclass (const char *bg=NULL)
	set background class (which will be ignored).
bool	set_mask_class (const char *mask)
void	set_silent ()
	set detector to silent: do not print results on std output
void	set_min_resolution (uint min_size)
void	set_max_resolution (uint max_size)
void	set_raw_thresholds (vector< float > &t)
	Set different thresholds for each scale during raw extraction.
void	set_nms (t_nms type=nms_overlap, float pre_threshold=0.0, float post_threshold=0.0, float pre_hfact=1.0, float pre_wfact=1.0, float post_hfact=1.0, float post_wfact=1.0, float woverh=1.0, float max_overlap=1.0, float max_hcenter_dist=0.0, float max_wcenter_dist=0.0, float vote_max_overlap=1.0, float vote_max_hcenter_dist=0.0, float vote_max_wcenter_dist=0.0)
void	set_scaler_mode (bool set)
	Enable or disable scaler mode, i.e. using scale prediction for boxes.
void	set_smoothing (uint type)
	Set output smoothing type. 0: none, 1: 3x3 kernel.
void	set_mem_optimization (Tstate &in, Tstate &out, bool keep_inputs=false)
void	set_netdim (idxdim &d)
	Set by hand the minimum network input and fix it.
void	set_outputs_dumping (const char *name)
void	set_bboxes_off ()
	Turn off the extraction of bounding boxes.
vector< string > &	get_labels ()
	Returns the vector of label strings.
void	set_ignore_outsiders ()
	Ignore bboxes that overlap with the outside of the image.
void	set_corners_inference (uint type)
void	set_bbox_decision (uint type)
void	set_bbox_scalings (mfidxdim &scalings)
template<class Tin >
bboxes &	fprop (idx< Tin > &img, const char *fname=NULL)
void	fprop_nms (bboxes &in, bboxes &out)
	Run non-maximum suppression on 'in' and put result in 'out'.
vector< idx< T > > &	get_originals ()
midx< T >	get_preprocessed (const bbox &b)
svector< midx< T > > &	get_preprocessed (bboxes &out, uint n=0, bool diverse=false, uint pre_diverse_max=100)
svector< midx< T > > &	get_preprocessed (bboxes &in, bboxes &out, uint n=0, bool diverse=false, uint pre_diverse_max=100)
idx< T >	get_mask (string &classname)
uint	get_total_saved ()
	Returns the number of bboxes saved so far.
string &	set_save (const string &directory, uint nmax=0, bool diversity=false)
void	init (idxdim &dinput, const char *frame_name=NULL)
	initialize dimensions and multi-resolution buffers.
Protected Member Functions
void	compute_scales (midxdim &scales, idxdim &netdim, idxdim &mindim, idxdim &maxdim, idxdim &indim, t_scaling type, uint nscales, double scales_step, const char *frame_name=NULL)
void	compute_resolutions (midxdim &scales, idxdim &mindim, idxdim &maxdim, uint nscales)
void	compute_resolutions (midxdim &scales, idxdim &indim, vector< double > &scale_factors)
void	compute_resolutions (midxdim &scales, idxdim &mindim, idxdim &maxdim, double scales_step)
void	compute_resolutions_up (midxdim &scales, idxdim &indim, idxdim &mindim, idxdim &maxdim, double scales_step)
void	validate_resolutions ()
void	smooth_outputs ()
	Smooth outputs.
void	update_merge_alignment ()
void	get_corners (mstate< Tstate > &outputs)
void	extract_bboxes (T threshold, bboxes &bboxes)
void	save_bboxes (bboxes &bboxes, const string &dir, const char *frame_name=NULL)
void	add_class (const char *name)
template<class Tin >
void	prepare (idx< Tin > &img, const char *fname=NULL)
void	prepare_scale (uint i)
void	multi_res_fprop ()
	do a fprop on thenet with multiple rescaled inputs
Protected Attributes
module_1_1< T, Tstate > &	thenet
	The network.
resizepp_module< T, Tstate > *	resizepp
	Resize module for multi-scaling.
bool	resizepp_delete
	We are responsible for deleting.
idx< T >	image
double	contrast
double	brightness
idx< float >	sizes
fstate_idx< T >	finput
Tstate *	input
	A forward buffer containing input image.
mstate< Tstate >	output
	output buffer
Tstate *	tmp
	tmp.
Tstate *	minput
	input buffer, used with mem optim.
svector< mstate< Tstate > >	ppinputs
	Preprocessed inputs of all scales.
svector< mstate< Tstate > >	outputs
	Output buffers of all scales.
vector< string >	labels
	String label of each class.
idxdim	indim
	Input dimensions.
idxdim	netdim
	network's input dimensions
bool	netdim_fixed
	Do not update netdim if true.
vector< rect< int > >	original_bboxes
	Bboxes in image after resizing.
int	bgclass
int	mask_class
idx< T >	mask
nms *	pnms
	Non-maximum suppression object.
midxdim	scales
	Multi-scale (ideal) scales.
midxdim	actual_scales
	Actually used scales.
midxdim	manual_scales
	Scales set manually.
vector< double >	scale_factors
	A list of scale factors.
uint	nscales
	Number of scales if set by hand.
double	scales_step
double	min_scale
	Minimum scale as factor of original res.
double	max_scale
	Maximum scale as factor of original res.
t_scaling	restype
	resolution type
bool	silent
	print results on std output if not silent
bool	save_mode
	save detected windows or not
string	save_dir
	directory where to save detections
vector< uint >	save_counts
	file counter for each class
bboxes	raw_bboxes
	raw bboxes extracted from outputs
bboxes	pruned_bboxes
	scale-pruned bboxes
uint	min_size
	minimum input size to network
uint	max_size
	maximum input size to network
vector< idx< T > >	odetections
	original windows yielding detection
svector< midx< T > >	ppdetections
	preprocessed wins yielding detection
bool	bodetections
	odetections is up-to-date or not
bool	bppdetections
	ppdetections is up-to-date or not
uint	save_max_per_frame
	max number of region saved
bool	diverse_ordering
	Saved samples diverse ordering.
bool	mem_optimization
	optimize memory or not.
bool	optimization_swap
	swap buffers or not.
bool	keep_inputs
uint	hzpad
	optimize input buffers or not.
uint	wzpad
	Zero-pad on height (each side).
std::ostream &	mout
	Zero-pad on width (each side).
std::ostream &	merr
	output stream.
uint	smoothing_type
	error output stream.
idx< T >	smoothing_kernel
bool	initialized
string	outputs_dump
	Outputs dumping name.
bool	bboxes_off
	Do not extract bboxes if true.
bool	adapt_scales
	Adapt scales to network structure.
bool	scaler_mode
answer_module< T, T, T, Tstate > *	answer
mstate< Tstate >	answers
	Buffers holding last answers.
bool	ignore_outsiders
	Ignore bbs overlapping outside.
uint	corners_inference
	0: from net 1: from net + save 2: load
bool	corners_infered
	Allows to infer only once.
mfidxdim	itl
mfidxdim	itr
mfidxdim	ibl
mfidxdim	ibr
	4 corners in input space.
mfidxdim	pptl
mfidxdim	pptr
mfidxdim	ppbl
mfidxdim	ppbr
	4 corners in pp input space.
float	pre_threshold
	Threshold for initial bbox extraction.
vector< float >	raw_thresholds
	Thresholds for each scale.
vector< uint >	scale_indices
	Input scales indices for each output.
uint	bbox_decision
	Decision type, 0: regular, 1: corners only.
mfidxdim	bbox_scalings
Friends
class	detector_gui
class	detection_thread
class	bootstrapping

template<typename T, class Tstate = fstate_idx<T>>
class ebl::detector< T, Tstate >

---

Constructor & Destructor Documentation

template<typename T , class Tstate >

ebl::detector< T, Tstate >::detector	(	module_1_1< T, Tstate > &	thenet,
		vector< string > &	labels,
		answer_module< T, T, T, Tstate > *	answer = NULL,
		resizepp_module< T, Tstate > *	resize = NULL,
		const char *	background = NULL,
		std::ostream &	out = std::cout,
		std::ostream &	err = std::cerr,
		bool	adapt_scales = false
	)

Constructor. Default resolutions are 1, 2 and 4 times the network's size. Resolutions can be set using set_resolutions(). Background class name default "bg" will be searched in the list of class names. To specify another background class, pass a non NULL background parameter.

Parameters:

labels	A vector of label strings.
pp	An optional resizing (and preprocessing) module, e.g. resizepp_module. If null, use resize_module by default.
background	The name of the background class. Default is "bg". If given, positive answers for this class are ignored.
adapt_scales	If true, adapt each scale so that they are valid input sizes for 'thenet' network. Otherwise, the network must crop inputs itself (see 'crop' attribute of modules).

---

Member Function Documentation

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::add_class

(

const char *

name

)

[protected]

Add a name to the vector of class names. This can be useful when generating intermediate classes from existing classes.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::compute_resolutions	(	midxdim &	scales,
		idxdim &	indim,
		vector< double > &	scale_factors
	)		[protected]

Compute each scale as a factor of 'indim' for each element of 'scale_factors' and put them into 'scales' vector.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::compute_resolutions	(	midxdim &	scales,
		idxdim &	mindim,
		idxdim &	maxdim,
		double	scales_step
	)		[protected]

Compute each scale with a step of 'scales_step' starting from 'maxdim' down to 'mindim'.

Parameters:

mindim	The minimum scale size.
maxdim	The maximum scale size.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::compute_resolutions	(	midxdim &	scales,
		idxdim &	mindim,
		idxdim &	maxdim,
		uint	nscales
	)		[protected]

Compute 'nscales' scales between 'mindim' and 'maxdim' resolutions and push them into 'scales' vector.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::compute_resolutions_up	(	midxdim &	scales,
		idxdim &	indim,
		idxdim &	mindim,
		idxdim &	maxdim,
		double	scales_step
	)		[protected]

Compute each scale with a step of 'scales_step' starting from 'mindim' up to 'maxdim'.

Parameters:

mindim	The minimum scale size.
maxdim	The maximum scale size.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::compute_scales	(	midxdim &	scales,
		idxdim &	netdim,
		idxdim &	mindim,
		idxdim &	maxdim,
		idxdim &	indim,
		t_scaling	type,
		uint	nscales,
		double	scales_step,
		const char *	frame_name = NULL
	)		[protected]

Compute all scales based on minimum, maximum and input dimensions, and scaling type.

Parameters:

netdim	The network's minimal input size.
mindim	The minimum scale size.
maxdim	The maximum scale size.
indim	The original input dimensions.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::extract_bboxes	(	T	threshold,
		bboxes &	bboxes
	)		[protected]

Extract bounding boxes with higher confidence than 'threshold' from internal 'outputs' into 'bboxes'.

template<typename T , class Tstate >

template<class Tin >

bboxes & ebl::detector< T, Tstate >::fprop	(	idx< Tin > &	img,
		const char *	fname = NULL
	)

fprop input image throught network. if image's and network's type differ, cast image into network's type through an idx_copy (avoid for better performance).

Parameters:

fname

Optional name for the frame being processed, used in the output files to be saved.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::get_corners

(

mstate< Tstate > &

outputs

)

[protected]

Fills internal buffers with 4 image corners coordinates in input and preprocessing space given the output sizes of 'outputs'.

template<typename T , class Tstate >

idx< T > ebl::detector< T, Tstate >::get_mask

(

string &

classname

)

Return a mask of output maps with the same size as the input. The mask is a max of all output resolution.

template<typename T , class Tstate >

vector< idx< T > > & ebl::detector< T, Tstate >::get_originals

(

)

Return a reference to a vector of windows in the original image that yielded a detection.

template<typename T , class Tstate >

svector< midx< T > > & ebl::detector< T, Tstate >::get_preprocessed	(	bboxes &	out,
		uint	n = 0,
		bool	diverse = false,
		uint	pre_diverse_max = 100
	)

Return a reference to a vector of windows in the preprocessed/scaled image that yielded a detection.

Parameters:

out	A vector of boxes filled with returned samples boxes.
n	Limit number of samples to 'n'. If n equals 0, return all.
diverse	If true, order samples by diversity.
pre_diverse_max	Limit the number of samples if diverse is enabled because it is an expensive process with complexity O(n^2).

template<typename T , class Tstate >

svector< midx< T > > & ebl::detector< T, Tstate >::get_preprocessed	(	bboxes &	in,
		bboxes &	out,
		uint	n = 0,
		bool	diverse = false,
		uint	pre_diverse_max = 100
	)

Return a reference to a vector of windows in the preprocessed/scaled image that yielded a detection.

Parameters:

in	Input boxes of samples to return.
out	A vector of boxes filled with returned samples boxes.
n	Limit number of samples to 'n'. If n equals 0, return all.
diverse	If true, order samples by diversity.
pre_diverse_max	Limit the number of samples if diverse is enabled because it is an expensive process with complexity O(n^2).

template<typename T , class Tstate >

midx< T > ebl::detector< T, Tstate >::get_preprocessed

(

const bbox &

b

)

Return the preprocessed input corresponding to bounding box 'b' or throws an exception if out of bounds.

template<typename T , class Tstate >

template<class Tin >

void ebl::detector< T, Tstate >::prepare	(	idx< Tin > &	img,
		const char *	fname = NULL
	)		[protected]

Prepare image and resolutions. This should be called before preprocess_resolution(). This mostly involves casting image into network's type and computing each scale's dimensions (no resizing) based on image's size.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::prepare_scale

(

uint

i

)

[protected]

Do preprocessing (resizing and channel/edge processing) for a particular resolution. This will set 'input' and 'output' buffers, that can then be used to fprop the network. This uses the 'image' member prepared by prepare() and should therefore be called after prepare().

Parameters:

res	The resolution to be preprocessed.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::save_bboxes	(	bboxes &	bboxes,
		const string &	dir,
		const char *	frame_name = NULL
	)		[protected]

save all bounding boxes of original (in original resolution) and preprocessed (resized and filtered) input into directory dir.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::set_bbox_decision

(

uint

type

)

Select the bbox extraction decision type. 0: decision based on confidence threshold. 1: only extract output corners.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::set_corners_inference

(

uint

type

)

The image corners transformation from outputs to input is infered back through the network (type is 0 or 1), and saved into "corners.mat" (type 1), or loaded from "corners.mat" only (type 2).

template<typename T , class Tstate >

bool ebl::detector< T, Tstate >::set_mask_class

(

const char *

mask

)

Set the mask class, which is ignored by bounding box detection. Instead, call get_mask() to retrieve a mask map of values above a given threshold. This is useful for continuous classes rather than discrete classes.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::set_max_resolution

(

uint

max_size

)

Set the maximum size of each side of an input to use as input to the network. I.e. a input to the network will be at most max_size * max_size big.

Parameters:

max_size

The maximum width or height in pixels.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::set_mem_optimization	(	Tstate &	in,
		Tstate &	out,
		bool	keep_inputs = false
	)

Enable memory optimization by using only 2 buffers (in and out) for entire flow. Those same buffers must have been passed to the network's constructor.

Parameters:

keep_inputs

If false, re-use input buffers for optimization. If true, some operations may be unavailable, such as saving the pre-processed detected windows.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::set_min_resolution

(

uint

min_size

)

Set the minimum size of each side of an input to use as input to the network. I.e. a input to the network will be at least min_size * min_size big.

Parameters:

min_size

The minimum width or height in pixels.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::set_nms	(	t_nms	type = nms_overlap,
		float	pre_threshold = 0.0,
		float	post_threshold = 0.0,
		float	pre_hfact = 1.0,
		float	pre_wfact = 1.0,
		float	post_hfact = 1.0,
		float	post_wfact = 1.0,
		float	woverh = 1.0,
		float	max_overlap = 1.0,
		float	max_hcenter_dist = 0.0,
		float	max_wcenter_dist = 0.0,
		float	vote_max_overlap = 1.0,
		float	vote_max_hcenter_dist = 0.0,
		float	vote_max_wcenter_dist = 0.0
	)

Enable nms of type 'type'. Refer to t_pruning declaration for different types. Default type is 1, regular pruning.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::set_outputs_dumping

(

const char *

name

)

Enables dumping of all outputs using the base name 'name', to which is appending the idx's size and '.mat'. Each resolution will be dump as a separate matrix file. Dumping will be called at the end of each fprop call for each resolution.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::set_resolutions	(	double	scales_steps,
		double	max_scale = 1.0,
		double	min_scale = 1.0
	)

Specify resolutions by the factor step, starting from factor 1 (network's size), adding factor_steps until reaching the original resolution.

Parameters:

max_scale	The maximum scale factor of the original resolution, 1.0 by default, i.e. the original resolution.
min_scale	The minimum scale factor of the smallest network size 1.0 by default, meaning the minimum network input size.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::set_resolutions

(

int

resolutions

)

Use nresolutions resolutions between the maximum resolution and the minimum resolution.

Parameters:

nresolutions

The number of resolutions to use.

template<typename T , class Tstate = fstate_idx<T>>

void ebl::detector< T, Tstate >::set_resolutions

(

idx< uint > &

resolutions

)

Specify resolutions by hand in an nx2 idx (heightxwidth), e.g. 240x320, 120x160.

Parameters:

resolutions

A uint idx containing resolutions (of size nx2)

template<typename T , class Tstate >

string & ebl::detector< T, Tstate >::set_save	(	const string &	directory,
		uint	nmax = 0,
		bool	diversity = false
	)

Enable saving of each (preprocessed) window inducing a positive detection into directory. All detections except for the background class are dumped into a directory corresponding to the class' name. This returns the directory string used.

Parameters:

nmax	Limit the number of windows saved per frame.
diversity	If true, order samples to be saved by diversity.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::set_zpads	(	float	hzpad,
		float	wzpad
	)

Add zero padding of (hzpad * the network's minimum input height) on each vertical sides and (wzpad * min width) on horizontal sides.

template<typename T , class Tstate >

void ebl::detector< T, Tstate >::update_merge_alignment

(

)

[protected]

If a merge module was found in the network, update its parameters so that merging is aligned on top left corner of all inputs.

template<typename T , class Tstate = fstate_idx<T>>

void ebl::detector< T, Tstate >::validate_resolutions

(

)

[protected]

checks that resolutions match the network size, if not adjust them. this method assumes nresolutions and resolutions members have already been initialized.

---

Member Data Documentation

template<typename T , class Tstate = fstate_idx<T>>

Tstate* ebl::detector< T, Tstate >::input [protected]

A forward buffer containing input image.

input buffer

---

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

• /home/rex/ebltrunk/core/libeblearn/include/detector.h
• /home/rex/ebltrunk/core/libeblearn/include/detector.hpp