added LogPolar Blind Spot Model (thanks to Fabio Solari for the contribution)

pull/13383/head
Vadim Pisarevsky 13 years ago
parent d10616775b
commit c8e206c2ab
  1. 214
      modules/contrib/include/opencv2/contrib/contrib.hpp
  2. 652
      modules/contrib/src/logpolar_bsm.cpp
  3. 82
      samples/cpp/logpolar_bsm.cpp

@ -44,6 +44,7 @@
#define __OPENCV_CONTRIB_HPP__
#include "opencv2/core/core.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/features2d/features2d.hpp"
#include "opencv2/objdetect/objdetect.hpp"
@ -633,14 +634,219 @@ namespace cv
TRANSLATION = 2,
RIGID_BODY_MOTION = 4
};
CV_EXPORTS bool RGBDOdometry( cv::Mat& Rt, const Mat& initRt,
const cv::Mat& image0, const cv::Mat& depth0, const cv::Mat& mask0,
const cv::Mat& image1, const cv::Mat& depth1, const cv::Mat& mask1,
const cv::Mat& cameraMatrix, float minDepth, float maxDepth, float maxDepthDiff,
CV_EXPORTS bool RGBDOdometry( Mat& Rt, const Mat& initRt,
const Mat& image0, const Mat& depth0, const Mat& mask0,
const Mat& image1, const Mat& depth1, const Mat& mask1,
const Mat& cameraMatrix, float minDepth, float maxDepth, float maxDepthDiff,
const std::vector<int>& iterCounts, const std::vector<float>& minGradientMagnitudes,
int transformType=RIGID_BODY_MOTION );
/**
*Bilinear interpolation technique.
*
*The value of a desired cortical pixel is obtained through a bilinear interpolation of the values
*of the four nearest neighbouring Cartesian pixels to the center of the RF.
*The same principle is applied to the inverse transformation.
*
*More details can be found in http://dx.doi.org/10.1007/978-3-642-23968-7_5
*/
class CV_EXPORTS LogPolar_Interp
{
public:
LogPolar_Interp() {}
/**
*Constructor
*\param w the width of the input image
*\param h the height of the input image
*\param center the transformation center: where the output precision is maximal
*\param R the number of rings of the cortical image (default value 70 pixel)
*\param ro0 the radius of the blind spot (default value 3 pixel)
*\param full \a 1 (default value) means that the retinal image (the inverse transform) is computed within the circumscribing circle.
* \a 0 means that the retinal image is computed within the inscribed circle.
*\param S the number of sectors of the cortical image (default value 70 pixel).
* Its value is usually internally computed to obtain a pixel aspect ratio equals to 1.
*\param sp \a 1 (default value) means that the parameter \a S is internally computed.
* \a 0 means that the parameter \a S is provided by the user.
*/
LogPolar_Interp(int w, int h, Point2i center, int R=70, double ro0=3.0,
int interp=INTER_LINEAR, int full=1, int S=117, int sp=1);
/**
*Transformation from Cartesian image to cortical (log-polar) image.
*\param source the Cartesian image
*\return the transformed image (cortical image)
*/
const Mat to_cortical(const Mat &source);
/**
*Transformation from cortical image to retinal (inverse log-polar) image.
*\param source the cortical image
*\return the transformed image (retinal image)
*/
const Mat to_cartesian(const Mat &source);
/**
*Destructor
*/
~LogPolar_Interp();
protected:
Mat Rsri;
Mat Csri;
int S, R, M, N;
int top, bottom,left,right;
double ro0, romax, a, q;
int interp;
Mat ETAyx;
Mat CSIyx;
void create_map(int M, int N, int R, int S, double ro0);
};
/**
*Overlapping circular receptive fields technique
*
*The Cartesian plane is divided in two regions: the fovea and the periphery.
*The fovea (oversampling) is handled by using the bilinear interpolation technique described above, whereas in
*the periphery we use the overlapping Gaussian circular RFs.
*
*More details can be found in http://dx.doi.org/10.1007/978-3-642-23968-7_5
*/
class CV_EXPORTS LogPolar_Overlapping
{
public:
LogPolar_Overlapping() {}
/**
*Constructor
*\param w the width of the input image
*\param h the height of the input image
*\param center the transformation center: where the output precision is maximal
*\param R the number of rings of the cortical image (default value 70 pixel)
*\param ro0 the radius of the blind spot (default value 3 pixel)
*\param full \a 1 (default value) means that the retinal image (the inverse transform) is computed within the circumscribing circle.
* \a 0 means that the retinal image is computed within the inscribed circle.
*\param S the number of sectors of the cortical image (default value 70 pixel).
* Its value is usually internally computed to obtain a pixel aspect ratio equals to 1.
*\param sp \a 1 (default value) means that the parameter \a S is internally computed.
* \a 0 means that the parameter \a S is provided by the user.
*/
LogPolar_Overlapping(int w, int h, Point2i center, int R=70,
double ro0=3.0, int full=1, int S=117, int sp=1);
/**
*Transformation from Cartesian image to cortical (log-polar) image.
*\param source the Cartesian image
*\return the transformed image (cortical image)
*/
const Mat to_cortical(const Mat &source);
/**
*Transformation from cortical image to retinal (inverse log-polar) image.
*\param source the cortical image
*\return the transformed image (retinal image)
*/
const Mat to_cartesian(const Mat &source);
/**
*Destructor
*/
~LogPolar_Overlapping();
protected:
Mat Rsri;
Mat Csri;
vector<int> Rsr;
vector<int> Csr;
vector<double> Wsr;
int S, R, M, N, ind1;
int top, bottom,left,right;
double ro0, romax, a, q;
struct kernel
{
kernel() { w = 0; }
vector<double> weights;
int w;
};
Mat ETAyx;
Mat CSIyx;
vector<kernel> w_ker_2D;
void create_map(int M, int N, int R, int S, double ro0);
};
/**
* Adjacent receptive fields technique
*
*All the Cartesian pixels, whose coordinates in the cortical domain share the same integer part, are assigned to the same RF.
*The precision of the boundaries of the RF can be improved by breaking each pixel into subpixels and assigning each of them to the correct RF.
*This technique is implemented from: Traver, V., Pla, F.: Log-polar mapping template design: From task-level requirements
*to geometry parameters. Image Vision Comput. 26(10) (2008) 1354-1370
*
*More details can be found in http://dx.doi.org/10.1007/978-3-642-23968-7_5
*/
class CV_EXPORTS LogPolar_Adjacent
{
public:
LogPolar_Adjacent() {}
/**
*Constructor
*\param w the width of the input image
*\param h the height of the input image
*\param center the transformation center: where the output precision is maximal
*\param R the number of rings of the cortical image (default value 70 pixel)
*\param ro0 the radius of the blind spot (default value 3 pixel)
*\param smin the size of the subpixel (default value 0.25 pixel)
*\param full \a 1 (default value) means that the retinal image (the inverse transform) is computed within the circumscribing circle.
* \a 0 means that the retinal image is computed within the inscribed circle.
*\param S the number of sectors of the cortical image (default value 70 pixel).
* Its value is usually internally computed to obtain a pixel aspect ratio equals to 1.
*\param sp \a 1 (default value) means that the parameter \a S is internally computed.
* \a 0 means that the parameter \a S is provided by the user.
*/
LogPolar_Adjacent(int w, int h, Point2i center, int R=70, double ro0=3.0, double smin=0.25, int full=1, int S=117, int sp=1);
/**
*Transformation from Cartesian image to cortical (log-polar) image.
*\param source the Cartesian image
*\return the transformed image (cortical image)
*/
const Mat to_cortical(const Mat &source);
/**
*Transformation from cortical image to retinal (inverse log-polar) image.
*\param source the cortical image
*\return the transformed image (retinal image)
*/
const Mat to_cartesian(const Mat &source);
/**
*Destructor
*/
~LogPolar_Adjacent();
protected:
struct pixel
{
pixel() { u = v = 0; a = 0.; }
int u;
int v;
double a;
};
int S, R, M, N;
int top, bottom,left,right;
double ro0, romax, a, q;
vector<vector<pixel> > L;
vector<double> A;
void subdivide_recursively(double x, double y, int i, int j, double length, double smin);
bool get_uv(double x, double y, int&u, int&v);
void create_map(int M, int N, int R, int S, double ro0, double smin);
};
}
#include "opencv2/contrib/retina.hpp"
#endif

@ -0,0 +1,652 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2012, Willow Garage Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The names of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
/*************************************************************************************
The LogPolar Blind Spot Model code has been contributed by Fabio Solari.
More details can be found in:
M. Chessa, S. P. Sabatini, F. Solari and F. Tatti (2011)
A Quantitative Comparison of Speed and Reliability for Log-Polar Mapping Techniques,
Computer Vision Systems - 8th International Conference,
ICVS 2011, Sophia Antipolis, France, September 20-22, 2011
(http://dx.doi.org/10.1007/978-3-642-23968-7_5)
***************************************************************************************/
#include "precomp.hpp"
#include <cmath>
#include <vector>
namespace cv
{
//------------------------------------interp-------------------------------------------
LogPolar_Interp::LogPolar_Interp(int w, int h, Point2i center, int R, double ro0, int interp, int full, int S, int sp)
{
if ( (center.x!=w/2 || center.y!=h/2) && full==0) full=1;
if (center.x<0) center.x=0;
if (center.y<0) center.y=0;
if (center.x>=w) center.x=w-1;
if (center.y>=h) center.y=h-1;
if (full){
int rtmp;
if (center.x<=w/2 && center.y>=h/2)
rtmp=(int)sqrt((float)center.y*center.y + (float)(w-center.x)*(w-center.x));
if (center.x>=w/2 && center.y>=h/2)
rtmp=(int)sqrt((float)center.y*center.y + (float)center.x*center.x);
if (center.x>=w/2 && center.y<=h/2)
rtmp=(int)sqrt((float)(h-center.y)*(h-center.y) + (float)center.x*center.x);
if (center.x<=w/2 && center.y<=h/2)
rtmp=(int)sqrt((float)(h-center.y)*(h-center.y) + (float)(w-center.x)*(w-center.x));
M=2*rtmp; N=2*rtmp;
top = M/2 - center.y;
bottom = M/2 - (h-center.y);
left = M/2 - center.x;
right = M/2 - (w - center.x);
}else{
top=bottom=left=right=0;
M=w; N=h;
}
if (sp){
int jc=M/2-1, ic=N/2-1;
int romax=min(ic, jc);
double a=exp(log((double)(romax/2-1)/(double)ro0)/(double)R);
S=(int) floor(2*M_PI/(a-1)+0.5);
}
this->interp=interp;
create_map(M, N, R, S, ro0);
}
void LogPolar_Interp::create_map(int M, int N, int R, int S, double ro0)
{
this->M=M;
this->N=N;
this->R=R;
this->S=S;
this->ro0=ro0;
int jc=N/2-1, ic=M/2-1;
romax=min(ic, jc);
a=exp(log((double)romax/(double)ro0)/(double)R);
q=((double)S)/(2*M_PI);
Rsri = Mat::zeros(S,R,CV_32FC1);
Csri = Mat::zeros(S,R,CV_32FC1);
ETAyx = Mat::zeros(N,M,CV_32FC1);
CSIyx = Mat::zeros(N,M,CV_32FC1);
for(int v=0; v<S; v++)
{
for(int u=0; u<R; u++)
{
Rsri.at<float>(v,u)=(float)(ro0*pow(a,u)*sin(v/q)+jc);
Csri.at<float>(v,u)=(float)(ro0*pow(a,u)*cos(v/q)+ic);
}
}
for(int j=0; j<N; j++)
{
for(int i=0; i<M; i++)
{
double theta;
if(i>=ic)
theta=atan((double)(j-jc)/(double)(i-ic));
else
theta=atan((double)(j-jc)/(double)(i-ic))+M_PI;
if(theta<0)
theta+=2*M_PI;
ETAyx.at<float>(j,i)=(float)(q*theta);
double ro2=(j-jc)*(j-jc)+(i-ic)*(i-ic);
CSIyx.at<float>(j,i)=(float)(0.5*log(ro2/(ro0*ro0))/log(a));
}
}
}
const Mat LogPolar_Interp::to_cortical(const Mat &source)
{
Mat out(S,R,CV_8UC1,Scalar(0));
Mat source_border;
copyMakeBorder(source,source_border,top,bottom,left,right,BORDER_CONSTANT,Scalar(0));
remap(source_border,out,Csri,Rsri,interp);
return out;
}
const Mat LogPolar_Interp::to_cartesian(const Mat &source)
{
Mat out(N,M,CV_8UC1,Scalar(0));
Mat source_border;
if (interp==INTER_NEAREST || interp==INTER_LINEAR){
copyMakeBorder(source,source_border,0,1,0,0,BORDER_CONSTANT,Scalar(0));
Mat rowS0 = source_border.row(S);
source_border.row(0).copyTo(rowS0);
} else if (interp==INTER_CUBIC){
copyMakeBorder(source,source_border,0,2,0,0,BORDER_CONSTANT,Scalar(0));
Mat rowS0 = source_border.row(S);
Mat rowS1 = source_border.row(S+1);
source_border.row(0).copyTo(rowS0);
source_border.row(1).copyTo(rowS1);
} else if (interp==INTER_LANCZOS4){
copyMakeBorder(source,source_border,0,4,0,0,BORDER_CONSTANT,Scalar(0));
Mat rowS0 = source_border.row(S);
Mat rowS1 = source_border.row(S+1);
Mat rowS2 = source_border.row(S+2);
Mat rowS3 = source_border.row(S+3);
source_border.row(0).copyTo(rowS0);
source_border.row(1).copyTo(rowS1);
source_border.row(2).copyTo(rowS2);
source_border.row(3).copyTo(rowS3);
}
remap(source_border,out,CSIyx,ETAyx,interp);
Mat out_cropped=out(Range(top,N-1-bottom),Range(left,M-1-right));
return out_cropped;
}
LogPolar_Interp::~LogPolar_Interp()
{
}
//------------------------------------overlapping----------------------------------
LogPolar_Overlapping::LogPolar_Overlapping(int w, int h, Point2i center, int R, double ro0, int full, int S, int sp)
{
if ( (center.x!=w/2 || center.y!=h/2) && full==0) full=1;
if (center.x<0) center.x=0;
if (center.y<0) center.y=0;
if (center.x>=w) center.x=w-1;
if (center.y>=h) center.y=h-1;
if (full){
int rtmp;
if (center.x<=w/2 && center.y>=h/2)
rtmp=(int)sqrt((float)center.y*center.y + (float)(w-center.x)*(w-center.x));
if (center.x>=w/2 && center.y>=h/2)
rtmp=(int)sqrt((float)center.y*center.y + (float)center.x*center.x);
if (center.x>=w/2 && center.y<=h/2)
rtmp=(int)sqrt((float)(h-center.y)*(h-center.y) + (float)center.x*center.x);
if (center.x<=w/2 && center.y<=h/2)
rtmp=(int)sqrt((float)(h-center.y)*(h-center.y) + (float)(w-center.x)*(w-center.x));
M=2*rtmp; N=2*rtmp;
top = M/2 - center.y;
bottom = M/2 - (h-center.y);
left = M/2 - center.x;
right = M/2 - (w - center.x);
}else{
top=bottom=left=right=0;
M=w; N=h;
}
if (sp){
int jc=M/2-1, ic=N/2-1;
int romax=min(ic, jc);
double a=exp(log((double)(romax/2-1)/(double)ro0)/(double)R);
S=(int) floor(2*M_PI/(a-1)+0.5);
}
create_map(M, N, R, S, ro0);
}
void LogPolar_Overlapping::create_map(int M, int N, int R, int S, double ro0)
{
this->M=M;
this->N=N;
this->R=R;
this->S=S;
this->ro0=ro0;
int jc=N/2-1, ic=M/2-1;
romax=min(ic, jc);
a=exp(log((double)romax/(double)ro0)/(double)R);
q=((double)S)/(2*M_PI);
ind1=0;
Rsri=Mat::zeros(S,R,CV_32FC1);
Csri=Mat::zeros(S,R,CV_32FC1);
ETAyx=Mat::zeros(N,M,CV_32FC1);
CSIyx=Mat::zeros(N,M,CV_32FC1);
Rsr.resize(R*S);
Csr.resize(R*S);
Wsr.resize(R);
w_ker_2D.resize(R*S);
for(int v=0; v<S; v++)
{
for(int u=0; u<R; u++)
{
Rsri.at<float>(v,u)=(float)(ro0*pow(a,u)*sin(v/q)+jc);
Csri.at<float>(v,u)=(float)(ro0*pow(a,u)*cos(v/q)+ic);
Rsr[v*R+u]=(int)floor(Rsri.at<float>(v,u));
Csr[v*R+u]=(int)floor(Csri.at<float>(v,u));
}
}
bool done=false;
for(int i=0; i<R; i++)
{
Wsr[i]=ro0*(a-1)*pow(a,i-1);
if((Wsr[i]>1)&&(done==false))
{
ind1=i;
done =true;
}
}
for(int j=0; j<N; j++)
{
for(int i=0; i<M; i++)//mdf
{
double theta;
if(i>=ic)
theta=atan((double)(j-jc)/(double)(i-ic));
else
theta=atan((double)(j-jc)/(double)(i-ic))+M_PI;
if(theta<0)
theta+=2*M_PI;
ETAyx.at<float>(j,i)=(float)(q*theta);
double ro2=(j-jc)*(j-jc)+(i-ic)*(i-ic);
CSIyx.at<float>(j,i)=(float)(0.5*log(ro2/(ro0*ro0))/log(a));
}
}
for(int v=0; v<S; v++)
for(int u=ind1; u<R; u++)
{
//double sigma=Wsr[u]/2.0;
double sigma=Wsr[u]/3.0;//modf
int w=(int) floor(3*sigma+0.5);
w_ker_2D[v*R+u].w=w;
w_ker_2D[v*R+u].weights.resize((2*w+1)*(2*w+1));
double dx=Csri.at<float>(v,u)-Csr[v*R+u];
double dy=Rsri.at<float>(v,u)-Rsr[v*R+u];
double tot=0;
for(int j=0; j<2*w+1; j++)
for(int i=0; i<2*w+1; i++)
{
(w_ker_2D[v*R+u].weights)[j*(2*w+1)+i]=exp(-(pow(i-w-dx, 2)+pow(j-w-dy, 2))/(2*sigma*sigma));
tot+=(w_ker_2D[v*R+u].weights)[j*(2*w+1)+i];
}
for(int j=0; j<(2*w+1); j++)
for(int i=0; i<(2*w+1); i++)
(w_ker_2D[v*R+u].weights)[j*(2*w+1)+i]/=tot;
}
}
const Mat LogPolar_Overlapping::to_cortical(const Mat &source)
{
Mat out(S,R,CV_8UC1,Scalar(0));
Mat source_border;
copyMakeBorder(source,source_border,top,bottom,left,right,BORDER_CONSTANT,Scalar(0));
remap(source_border,out,Csri,Rsri,INTER_LINEAR);
int wm=w_ker_2D[R-1].w;
vector<int> IMG((M+2*wm+1)*(N+2*wm+1), 0);
for(int j=0; j<N; j++)
for(int i=0; i<M; i++)
IMG[(M+2*wm+1)*(j+wm)+i+wm]=source_border.at<uchar>(j,i);
for(int v=0; v<S; v++)
for(int u=ind1; u<R; u++)
{
int w=w_ker_2D[v*R+u].w;
double tmp=0;
for(int rf=0; rf<(2*w+1); rf++)
{
for(int cf=0; cf<(2*w+1); cf++)
{
double weight=(w_ker_2D[v*R+u]).weights[rf*(2*w+1)+cf];
tmp+=IMG[(M+2*wm+1)*((rf-w)+Rsr[v*R+u]+wm)+((cf-w)+Csr[v*R+u]+wm)]*weight;
}
}
out.at<uchar>(v,u)=(uchar) floor(tmp+0.5);
}
return out;
}
const Mat LogPolar_Overlapping::to_cartesian(const Mat &source)
{
Mat out(N,M,CV_8UC1,Scalar(0));
Mat source_border;
copyMakeBorder(source,source_border,0,1,0,0,BORDER_CONSTANT,Scalar(0));
Mat rowS = source_border.row(S);
source_border.row(0).copyTo(rowS);
remap(source_border,out,CSIyx,ETAyx,INTER_LINEAR);
int wm=w_ker_2D[R-1].w;
vector<double> IMG((N+2*wm+1)*(M+2*wm+1), 0.);
vector<double> NOR((N+2*wm+1)*(M+2*wm+1), 0.);
for(int v=0; v<S; v++)
for(int u=ind1; u<R; u++)
{
int w=w_ker_2D[v*R+u].w;
for(int j=0; j<(2*w+1); j++)
{
for(int i=0; i<(2*w+1); i++)
{
int ind=(M+2*wm+1)*((j-w)+Rsr[v*R+u]+wm)+(i-w)+Csr[v*R+u]+wm;
IMG[ind]+=((w_ker_2D[v*R+u]).weights[j*(2*w+1)+i])*source.at<uchar>(v, u);
NOR[ind]+=((w_ker_2D[v*R+u]).weights[j*(2*w+1)+i]);
}
}
}
for(int i=0; i<((N+2*wm+1)*(M+2*wm+1)); i++)
IMG[i]/=NOR[i];
//int xc=M/2-1, yc=N/2-1;
for(int j=wm; j<N+wm; j++)
for(int i=wm; i<M+wm; i++)
{
/*if(NOR[(M+2*wm+1)*j+i]>0)
ret[M*(j-wm)+i-wm]=(int) floor(IMG[(M+2*wm+1)*j+i]+0.5);*/
//int ro=(int)floor(sqrt((double)((j-wm-yc)*(j-wm-yc)+(i-wm-xc)*(i-wm-xc))));
int csi=(int) floor(CSIyx.at<float>(j-wm,i-wm));
if((csi>=(ind1-(w_ker_2D[ind1]).w))&&(csi<R))
out.at<uchar>(j-wm,i-wm)=(uchar) floor(IMG[(M+2*wm+1)*j+i]+0.5);
}
Mat out_cropped=out(Range(top,N-1-bottom),Range(left,M-1-right));
return out_cropped;
}
LogPolar_Overlapping::~LogPolar_Overlapping()
{
}
//----------------------------------------adjacent---------------------------------------
LogPolar_Adjacent::LogPolar_Adjacent(int w, int h, Point2i center, int R, double ro0, double smin, int full, int S, int sp)
{
if ( (center.x!=w/2 || center.y!=h/2) && full==0) full=1;
if (center.x<0) center.x=0;
if (center.y<0) center.y=0;
if (center.x>=w) center.x=w-1;
if (center.y>=h) center.y=h-1;
if (full){
int rtmp;
if (center.x<=w/2 && center.y>=h/2)
rtmp=(int)sqrt((float)center.y*center.y + (float)(w-center.x)*(w-center.x));
if (center.x>=w/2 && center.y>=h/2)
rtmp=(int)sqrt((float)center.y*center.y + (float)center.x*center.x);
if (center.x>=w/2 && center.y<=h/2)
rtmp=(int)sqrt((float)(h-center.y)*(h-center.y) + (float)center.x*center.x);
if (center.x<=w/2 && center.y<=h/2)
rtmp=(int)sqrt((float)(h-center.y)*(h-center.y) + (float)(w-center.x)*(w-center.x));
M=2*rtmp; N=2*rtmp;
top = M/2 - center.y;
bottom = M/2 - (h-center.y);
left = M/2 - center.x;
right = M/2 - (w - center.x);
}else{
top=bottom=left=right=0;
M=w; N=h;
}
if (sp){
int jc=M/2-1, ic=N/2-1;
int romax=min(ic, jc);
double a=exp(log((double)(romax/2-1)/(double)ro0)/(double)R);
S=(int) floor(2*M_PI/(a-1)+0.5);
}
create_map(M, N, R, S, ro0, smin);
}
void LogPolar_Adjacent::create_map(int M, int N, int R, int S, double ro0, double smin)
{
LogPolar_Adjacent::M=M;
LogPolar_Adjacent::N=N;
LogPolar_Adjacent::R=R;
LogPolar_Adjacent::S=S;
LogPolar_Adjacent::ro0=ro0;
romax=min(M/2.0, N/2.0);
a=exp(log(romax/ro0)/(double)R);
q=S/(2*M_PI);
A.resize(R*S);
L.resize(M*N);
for(int i=0; i<R*S; i++)
A[i]=0;
double xc=M/2.0, yc=N/2.0;
for(int j=0; j<N; j++)
for(int i=0; i<M; i++)
{
double x=i+0.5-xc, y=j+0.5-yc;
subdivide_recursively(x, y, i, j, 1, smin);
}
}
void LogPolar_Adjacent::subdivide_recursively(double x, double y, int i, int j, double length, double smin)
{
if(length<=smin)
{
int u, v;
if(get_uv(x, y, u, v))
{
pixel p;
p.u=u;
p.v=v;
p.a=length*length;
L[M*j+i].push_back(p);
A[v*R+u]+=length*length;
}
}
if(length>smin)
{
double xs[4], ys[4];
int us[4], vs[4];
xs[0]=xs[3]=x+length/4.0;
xs[1]=xs[2]=x-length/4.0;
ys[1]=ys[0]=y+length/4.0;
ys[2]=ys[3]=y-length/4.0;
for(int z=0; z<4; z++)
get_uv(xs[z], ys[z], us[z], vs[z]);
bool c=true;
for(int w=1; w<4; w++)
{
if(us[w]!=us[w-1])
c=false;
if(vs[w]!=vs[w-1])
c=false;
}
if(c)
{
if(us[0]!=-1)
{
pixel p;
p.u=us[0];
p.v=vs[0];
p.a=length*length;
L[M*j+i].push_back(p);
A[vs[0]*R+us[0]]+=length*length;
}
}
else
{
for(int z=0; z<4; z++)
if(us[z]!=-1)
subdivide_recursively(xs[z], ys[z], i, j, length/2.0, smin);
}
}
}
const Mat LogPolar_Adjacent::to_cortical(const Mat &source)
{
Mat source_border;
copyMakeBorder(source,source_border,top,bottom,left,right,BORDER_CONSTANT,Scalar(0));
vector<double> map(R*S, 0.);
for(int j=0; j<N; j++)
for(int i=0; i<M; i++)
{
for(size_t z=0; z<(L[M*j+i]).size(); z++)
{
map[R*((L[M*j+i])[z].v)+((L[M*j+i])[z].u)]+=((L[M*j+i])[z].a)*(source_border.at<uchar>(j,i));
}
}
for(int i=0; i<R*S; i++)
map[i]/=A[i];
Mat out(S,R,CV_8UC1,Scalar(0));
for(int i=0; i<S; i++)
for(int j=0;j<R;j++)
out.at<uchar>(i,j)=(uchar) floor(map[i*R+j]+0.5);
return out;
}
const Mat LogPolar_Adjacent::to_cartesian(const Mat &source)
{
vector<double> map(M*N, 0.);
for(int j=0; j<N; j++)
for(int i=0; i<M; i++)
{
for(size_t z=0; z<(L[M*j+i]).size(); z++)
{
map[M*j+i]+=(L[M*j+i])[z].a*source.at<uchar>((L[M*j+i])[z].v,(L[M*j+i])[z].u);
}
}
Mat out(N,M,CV_8UC1,Scalar(0));
for(int i=0; i<N; i++)
for(int j=0; j<M; j++)
out.at<uchar>(i,j)=(uchar) floor(map[i*M+j]+0.5);
Mat out_cropped=out(Range(top,N-1-bottom),Range(left,M-1-right));
return out_cropped;
}
bool LogPolar_Adjacent::get_uv(double x, double y, int&u, int&v)
{
double ro=sqrt(x*x+y*y), theta;
if(x>0)
theta=atan(y/x);
else
theta=atan(y/x)+M_PI;
if(ro<ro0||ro>romax)
{
u=-1;
v=-1;
return false;
}
else
{
u= (int) floor(log(ro/ro0)/log(a));
if(theta>=0)
v= (int) floor(q*theta);
else
v= (int) floor(q*(theta+2*M_PI));
return true;
}
}
LogPolar_Adjacent::~LogPolar_Adjacent()
{
}
}

@ -0,0 +1,82 @@
/*Authors
* Manuela Chessa, Fabio Solari, Fabio Tatti, Silvio P. Sabatini
*
* manuela.chessa@unige.it, fabio.solari@unige.it
*
* PSPC-lab - University of Genoa
*/
#include "opencv2/opencv.hpp"
#include <iostream>
#include <cmath>
using namespace cv;
using namespace std;
void help()
{
cout << "LogPolar Blind Spot Model sample.\nShortcuts:"
"\n\tn for nearest pixel technique"
"\n\tb for bilinear interpolation technique"
"\n\to for overlapping circular receptive fields"
"\n\ta for adjacent receptive fields"
"\n\tq or ESC quit\n";
}
int main(int argc, char** argv)
{
Mat img = imread(argc > 1 ? argv[1] : "lena.jpg",1); // open the image
if(img.empty()) // check if we succeeded
{
cout << "can not load image\n";
return 0;
}
help();
Size s=img.size();
int w=s.width, h=s.height;
int ro0=3; //radius of the blind spot
int R=120; //number of rings
//Creation of the four different objects that implement the four log-polar transformations
//Off-line computation
Point2i center(w/2,h/2);
LogPolar_Interp nearest(w, h, center, R, ro0, INTER_NEAREST);
LogPolar_Interp bilin(w,h, center,R,ro0);
LogPolar_Overlapping overlap(w,h,center,R,ro0);
LogPolar_Adjacent adj(w,h,center,R,ro0,0.25);
namedWindow("Cartesian",1);
namedWindow("retinal",1);
namedWindow("cortical",1);
int wk='n';
Mat Cortical, Retinal;
//On-line computation
for(;;)
{
if(wk=='n'){
Cortical=nearest.to_cortical(img);
Retinal=nearest.to_cartesian(Cortical);
}else if (wk=='b'){
Cortical=bilin.to_cortical(img);
Retinal=bilin.to_cartesian(Cortical);
}else if (wk=='o'){
Cortical=overlap.to_cortical(img);
Retinal=overlap.to_cartesian(Cortical);
}else if (wk=='a'){
Cortical=adj.to_cortical(img);
Retinal=adj.to_cartesian(Cortical);
}
imshow("Cartesian", img);
imshow("cortical", Cortical);
imshow("retinal", Retinal);
int c=waitKey(15);
if (c>0) wk=c;
if(wk =='q' || (wk & 255) == 27) break;
}
return 0;
}
Loading…
Cancel
Save