icl::filter::IntegralImgOp Class Reference

class for creating integral images More...

#include <IntegralImgOp.h>

Inheritance diagram for icl::filter::IntegralImgOp:

List of all members.

Public Member Functions
	IntegralImgOp (core::depth integralImageDepth=core::depth32s)
	Constructor.
	~IntegralImgOp ()
	Destructor.
void	setIntegralImageDepth (core::depth integralImageDepth)
	sets the depth of the integralImage (depth8u etc)
core::depth	getIntegralImageDepth () const
	returns the depth of the integralImage
void	apply (const core::ImgBase *src, core::ImgBase **dst)
	applies the integralimage Operaor
Private Attributes
core::depth	m_integralImageDepth
	destination depth
core::ImgBase *	m_buf
	used only if IPP is available

---

Detailed Description

class for creating integral images

The integral image A of an image a (introduced by Viola & Jones (2001) in the paper : "Rapid Object Detection using a Boosted Cascade of Simple Features" ) defined as follows:

It can be calculated incrementally using the following equation:

illustration:

      +++++..
      +++CA..
      +++BX..
      .......
      X = src(x) + A + B - C
    

Implementation

Other definitions

There are alternative definitions (like the ipp):

From this it follows, that the first integral image row and colum is filled with zeros and the internal image grows by one pixel in each dimension, with the benefit of some better usability of the result image. We use the above definition, which is a bit more convenient for the most common applications in our sight.

IPP Optimization

IPP optimization is not used, as IPP uses a different formulation of the integral image which leads to an integral image that is one row and one column larger as the source image. IPPs integral image formulation is

. As our implementation comes really close to the IPP-version in terms of performance, we dont use the IPP for this UnaryOp.

Supported Type-Combinations

We support all source image depth, the integral image always needs a large value domain, so here, only icl32s, icl32f and icl64f are supported.

Performance

The calculation is very fast and we come close to the IPP performace.

• Benchmark plattworm: Intel Core2Duo with 2GHz, 2GB Ram (single core-performance)
• Test Image: 1000x1000, 1-channel, icl8u (uchar)
• Destination depth 32s (int)
• Compiler: g++ 4.3
• Compilation flags: -O4 -march=native
• Times
  • our implementation: 2.3ms
  • Intel IPP 1.8ms (but different integral image, and not supported)

---

Constructor & Destructor Documentation

icl::filter::IntegralImgOp::IntegralImgOp

(

core::depth

integralImageDepth = core::depth32s

)

Constructor.

Parameters:

integralImageDepth

the depth of the integralImage (depth8u etc)

icl::filter::IntegralImgOp::~IntegralImgOp

(

)

Destructor.

---

Member Function Documentation

void icl::filter::IntegralImgOp::apply	(	const core::ImgBase *	src,
		core::ImgBase **	dst
	)		[virtual]

applies the integralimage Operaor

Parameters:

src	The source image
dst	Pointer to the destination image

Implements icl::filter::UnaryOp.

core::depth icl::filter::IntegralImgOp::getIntegralImageDepth

(

)

const

returns the depth of the integralImage

Returns:

the depth of the integralImage

void icl::filter::IntegralImgOp::setIntegralImageDepth

(

core::depth

integralImageDepth

)

sets the depth of the integralImage (depth8u etc)

Parameters:

integralImageDepth

the depth of the integralImage (depth8u etc)

---

Member Data Documentation

core::ImgBase* icl::filter::IntegralImgOp::m_buf [private]

used only if IPP is available

Reimplemented from icl::filter::UnaryOp.

core::depth icl::filter::IntegralImgOp::m_integralImageDepth [private]

destination depth

---

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

• /home/jenkins/workspace/icl-manual-8.0/icl-manual/ICLFilter/src/ICLFilter/IntegralImgOp.h