Related
my c program is running to slow (right now it is around 40 seconds without parallelization). I have tried using openmp which has brought the timing down significantly but I am looking to use simple and natural ways to make my code run faster other than using parallel for loops. The basic structure of the code is that is takes some command line arguments as inputs and then saves those inputs as variables. Then it recursively computes a variable called Rplus1 using the math.h library and the complex.h library. The problem of the code and where it is taking most of it's time is at the bottom where there are nested for loops. My goal is to get the whole code running in under 5 seconds but as of now it runs in about 40 seconds without using parallel for loops. Please Help!
#include "time.h"
#include "stdio.h"
#include "stdlib.h"
#include "complex.h"
#include "math.h"
#include "string.h"
#include "unistd.h"
#include "omp.h"
#define PI 3.14159265
int main (int argc, char *argv[]){
if(argc >= 8){
double start1 = omp_get_wtime();
// command line arguments are aligned in the following order: [theta] [number of layers in superlattice] [material_1] [lat const_1] [number of unit cells_1] [material_2] [lat const_2] [number of unit cells_2] .... [material_N] [lat const_N] [number of unit cells_N] [Log/Linear] [number of repeating superlattice layers] [yes/no]
int N;
sscanf(argv[2],"%d",&N); // Number of layers in superlattice specified by second input argument
if(strcmp(argv[argc-1],"yes") == 0) //If the substrate is included then add one more layer to the N variable
{
N = N+1;
}
int total;
sscanf(argv[argc-2],"%d",&total); // Number of repeating superlattice layers specified by second to last argument
double layers[N][6], horizangle[1001], vertangle[1001];
double complex (*F_hkl)[1001][1001] = malloc(N*1001*1001*sizeof(complex double)), (*F_0)[1001][1001] = malloc(N*1001*1001*sizeof(complex double)), (*g)[1001][1001] = malloc(N*1001*1001*sizeof(complex double)), (*g_0)[1001][1001] = malloc(N*1001*1001*sizeof(complex double)),SF_table[10];// this array will hold the unit cell structure factors for all of the materials selected for each wavevector in the beam spectrum
double real, real2, lam, c_light = 299792458, h_pl = 4.135667516e-15,E = 10e3, r_0 = 2.818e-15, Lccd = 1.013;// just a few variables to hold values through calculations and constants, speed of light, plancks const, photon energy, and detector distance from sample
double angle;
double complex z;// just a variable to hold complex numbers throughout calculations
int i,j,m,n,t; // integers to index through arrays
lam = (h_pl*c_light)/E;
sscanf(argv[1],"%lf",&angle); //first argument is the angle of incidence, read it
angle = angle*(PI/180.0);
angle2 = -angle;
double (*table)[10] = malloc(10*9*sizeof(double)); // this array holds all the coefficients to calculate the atomic scattering factor below
double (*table2)[10] = malloc(10*2*sizeof(double));
FILE*datfile1 = fopen("/home/vhosts/xraydev.engr.wisc.edu/data/coef_table.bin","rb"); // read the binary file containg all the coefficients
fread(table,sizeof(double),90,datfile1);
fclose(datfile1);
FILE*datfile2 = fopen("/home/vhosts/xraydev.engr.wisc.edu/data/dispersioncs.bin","rb");
fread(table2,sizeof(double),20,datfile2);
fclose(datfile2);
// Calculate scattering factors for all elements
double a,b;
double k_z = (sin(angle)/lam)*1e-10; // incorporate angular dependence of SF but neglect 0.24 degree divergence because of approximation
for(i = 0;i<10;i++) // for each element...
{
SF_table[i] = 0;
for(j = 0;j<4;j++) // summation
{
a = table[2*j][i];
b = table[2*j+1][i];
SF_table[i] = SF_table[i] + a * exp(-b*k_z*k_z);
}
SF_table[i] = SF_table[i] + table[8][i] + table2[0][i] + table2[1][i]*I;
}
free(table);
double mm = 4.0, (*phi)[1001][1001] = malloc(N*1001*1001*sizeof(double));
for(i = 1; i < N+1; i++) // for each layer of material...
{
sscanf(argv[i*3+1],"%lf",&layers[i-1][1]); // get out of plane lattice constant
sscanf(argv[i*3+2],"%lf",&layers[i-1][2]); // get the number of unit cells in the layer
layers[i-1][1] = layers[i-1][1]*1e-10; // convert lat const input to meters
// Define reciprocal space positions at the incident angle h, k, l
layers[i-1][3] = 0; // h
layers[i-1][4] = 0; // k
double l; // l calculated for each wavevector in the spectrum because l changes with angle of incidence
for (m = 0; m < 1001; m++)
{
for (n = 0; n <1001; n++)
{
l = 4;
phi[i-1][m][n] = 2*PI*layers[i-1][1]*sin(angle)/lam; // Caculate phi for each layer
if(strcmp(argv[i*3],"GaAs") == 0)
{
F_hkl[i-1][m][n] = (2+2*cexp(I*PI*l))*(SF_table[2]+SF_table[3]*cexp(I*PI*l/2));
F_0[i-1][m][n] = 0.5*8.0*(31 + table2[0][2] + table2[1][2]*I) + 0.5*8.0*(33 + table2[0][3] + table2[1][3]*I);
g[i-1][m][n] = 2*r_0*F_hkl[i-1][m][n]/mm/layers[i-1][1]*cos(2*angle[m][n]);
g_0[i-1][m][n] = 2*r_0*F_0[i-1][m][n]/mm/layers[i-1][1];
}
if(strcmp(argv[i*3],"AlGaAs") == 0)
{
F_hkl[i-1][m][n] = (2+2*cexp(I*PI*l))*((0.76*SF_table[2]+ 0.24*SF_table[4])+SF_table[3]*cexp(I*PI*l/2));
F_0[i-1][m][n] = 0.24*4.0*(13 + table2[0][4] + table2[1][4]*I) + 0.76*4.0*(31 + table2[0][2] + table2[1][2]*I) + 4.0*(33 + table2[0][3] + table2[1][3]*I);
g[i-1][m][n] = 2*r_0*F_hkl[i-1][m][n]/mm/layers[i-1][1]*cos(2*angle[m][n]);
g_0[i-1][m][n] = 2*r_0*F_0[i-1][m][n]/mm/layers[i-1][1];
}
}
}
}
double complex (*Rplus1)[1001] = malloc(1001*1001*sizeof(double complex));
for (m = 0; m < 1001; m++)
{
for (n = 0; n <1001; n++)
{
Rplus1[m][n] = 0.0;
}
}
double stop1 = omp_get_wtime();
for(i=1;i<N;i++) // For each layer of the film
{
for(j=0;j<layers[i][2];j++) // For each unit cell
{
for (m = 0; m < 1001; m++) // For each row of the diffraction pattern
{
for (n = 0; n <1001; n++) // For each column of the diffraction pattern
{
Rplus1[m][n] = -I*g[i][m][n] + ((1-I*g_0[i][m][n])*(1-I*g_0[i][m][n]))/(I*g[i][m][n] + (cos(-2*phi[i][m][n])+I*sin(-2*phi[i][m][n]))/Rplus1[m][n]);
}
}
}
}
double stop2 = omp_get_wtime();
double elapsed1 = (double)(stop1 - start1);// Second user defined function to use Durbin and Follis recursive formula
double elapsed2 = (double)(stop2 - start1);// Second user defined function to use Durbin and Follis recursive formula
printf("main() through before diffraction function took %f seconds to run\n\n",elapsed1);
printf("main() through after diffraction function took %f seconds to run\n\n",elapsed2);
}
}
I've got an error, regarding calling JacobiSVD in my cuda function.
This is the part of the code that causing the error.
Eigen::JacobiSVD<Eigen::Matrix3d> svd( cov_e, Eigen::ComputeThinU | Eigen::ComputeThinV);
And this is the error message.
CUDA_voxel_building.cu(43): error: calling a __host__
function("Eigen::JacobiSVD , (int)2> ::JacobiSVD") from a __global__
function("kernel") is not allowed
I've used the following command to compile it.
nvcc -std=c++11 -D_MWAITXINTRIN_H_INCLUDED -D__STRICT_ANSI__ -ptx CUDA_voxel_building.cu
I'm using code 8.0 with eigen3 on ubuntu 16.04.
It seems like other functions such as eigen value decomposition also gives the same error.
Anyone knows a solution? I'm enclosing my code below.
//nvcc -ptx CUDA_voxel_building.cu
#include </usr/include/eigen3/Eigen/Core>
#include </usr/include/eigen3/Eigen/SVD>
/*
#include </usr/include/eigen3/Eigen/Sparse>
#include </usr/include/eigen3/Eigen/Dense>
#include </usr/include/eigen3/Eigen/Eigenvalues>
*/
__global__ void kernel(double *p, double *breaks,double *ind, double *mu, double *cov, double *e,double *v, int *n, char *isgood, int minpts, int maxgpu){
bool debuginfo = false;
int idx = threadIdx.x + blockIdx.x * blockDim.x;
if(debuginfo)printf("Thread %d got pointer\n",idx);
if( idx < maxgpu){
int s_ind = breaks[idx];
int e_ind = breaks[idx+1];
int diff = e_ind-s_ind;
if(diff >minpts){
int cnt = 0;
Eigen::MatrixXd local_p(3,diff) ;
for(int k = s_ind;k<e_ind;k++){
int temp_ind=ind[k];
//Eigen::Matrix<double, 3, diff> local_p;
local_p(1,cnt) = p[temp_ind*3];
local_p(2,cnt) = p[temp_ind*3+1];
local_p(3,cnt) = p[temp_ind*3+2];
cnt++;
}
Eigen::Matrix3d centered = local_p.rowwise() - local_p.colwise().mean();
Eigen::Matrix3d cov_e = (centered.adjoint() * centered) / double(local_p.rows() - 1);
Eigen::JacobiSVD<Eigen::Matrix3d> svd( cov_e, Eigen::ComputeThinU | Eigen::ComputeThinV);
/* Eigen::Matrix3d Cp = svd.matrixU() * svd.singularValues().asDiagonal() * svd.matrixV().transpose();
mu[idx]=p[ind[s_ind]*3];
mu[idx+1]=p[ind[s_ind+1]*3];
mu[idx+2]=p[ind[s_ind+2]*3];
e[idx]=svd.singularValues()(0);
e[idx+1]=svd.singularValues()(1);
e[idx+2]=svd.singularValues()(2);
n[idx] = diff;
isgood[idx] = 1;
for(int x = 0; x < 3; x++)
{
for(int y = 0; y < 3; y++)
{
v[x+ 3*y +idx*9]=svd.matrixV()(x, y);
cov[x+ 3*y +idx*9]=cov_e(x, y);
//if(debuginfo)printf("%f ",R[x+ 3*y +i*9]);
if(debuginfo)printf("%f ",Rm(x, y));
}
}
*/
} else {
mu[idx]=0;
mu[idx+1]=0;
mu[idx+2]=0;
e[idx]=0;
e[idx+1]=0;
e[idx+2]=0;
n[idx] = 0;
isgood[idx] = 0;
for(int x = 0; x < 3; x++)
{
for(int y = 0; y < 3; y++)
{
v[x+ 3*y +idx*9]=0;
cov[x+ 3*y +idx*9]=0;
}
}
}
}
}
First of all, Ubuntu 16.04 provides Eigen 3.3-beta1, which is not really recommended to be used. I would suggest upgrading to a more recent version. Furthermore, to include Eigen, write (e.g.):
#include <Eigen/Eigenvalues>
and compile with -I /usr/include/eigen3 (if you use the version provided by the OS), or better -I /path/to/local/eigen-version.
Then, as talonmies noted, you can't call host-functions from kernels, (I'm not sure at the moment, why JacobiSVD is not marked as device function), but in your case it would make much more sense to use Eigen::SelfAdjointEigenSolver, anyway. Since the matrix you are decomposing is fixed-size 3x3 you should actually use the optimized computeDirect method:
Eigen::SelfAdjointEigenSolver<Eigen::Matrix3d> eig; // default constructor
eig.computeDirect(cov_e); // works for 2x2 and 3x3 matrices, does not require loops
It seems the computeDirect even works on the beta version provided by Ubuntu (I'd still recommend to update).
Some unrelated notes:
The following is wrong, since you should start with index 0:
local_p(1,cnt) = p[temp_ind*3];
local_p(2,cnt) = p[temp_ind*3+1];
local_p(3,cnt) = p[temp_ind*3+2];
Also, you can write this in one line:
local_p.col(cnt) = Eigen::Vector3d::Map(p+temp_ind*3);
This line will not fit (unless diff==3):
Eigen::Matrix3d centered = local_p.rowwise() - local_p.colwise().mean();
What you probably mean is (local_p is actually 3xn not nx3)
Eigen::Matrix<double, 3, Eigen::Dynamic> centered = local_p.colwise() - local_p.rowwise().mean();
And when computing cov_e you need to .adjoint() the second factor, not the first.
You can avoid both 'big' matrices local_p and centered, by directly accumulating Eigen::Matrix3d sum2 and Eigen::Vector3d sum with sum2 += v*v.adjoint() and sum +=v and computing
Eigen::Vector3d mu = sum / diff;
Eigen::Matrix3d cov_e = (sum2 - mu*mu.adjoint()*diff)/(diff-1);
I need a data structure for storing float values at an uniformly sampled 3D mesh:
x = x0 + ix*dx where 0 <= ix < nx
y = y0 + iy*dy where 0 <= iy < ny
z = z0 + iz*dz where 0 <= iz < nz
Up to now I have used my Array class:
Array3D<float> A(nx, ny,nz);
A(0,0,0) = 0.0f; // ix = iy = iz = 0
Internally it stores the float values as an 1D array with nx * ny * nz elements.
However now I need to represent an mesh with more values than I have RAM,
e.g. nx = ny = nz = 2000.
I think many neighbour nodes in such an mesh may have similar values so I was thinking if there was some simple way that I could "coarsen" the mesh adaptively.
For instance if the 8 (ix,iy,iz) nodes of an cell in this mesh have values that are less than 5% apart; they are "removed" and replaced by just one value; the mean of the 8 values.
How could I implement such a data structure in a simple and efficient way?
EDIT:
thanks Ante for suggesting lossy compression. I think this could work the following way:
#define BLOCK_SIZE 64
struct CompressedArray3D {
CompressedArray3D(int ni, int nj, int nk) {
NI = ni/BLOCK_SIZE + 1;
NJ = nj/BLOCK_SIZE + 1;
NK = nk/BLOCK_SIZE + 1;
blocks = new float*[NI*NJ*NK];
compressedSize = new unsigned int[NI*NJ*NK];
}
void setBlock(int I, int J, int K, float values[BLOCK_SIZE][BLOCK_SIZE][BLOCK_SIZE]) {
unsigned int csize;
blocks[I*NJ*NK + J*NK + K] = compress(values, csize);
compressedSize[I*NJ*NK + J*NK + K] = csize;
}
float getValue(int i, int j, int k) {
int I = i/BLOCK_SIZE;
int J = j/BLOCK_SIZE;
int K = k/BLOCK_SIZE;
int ii = i - I*BLOCK_SIZE;
int jj = j - J*BLOCK_SIZE;
int kk = k - K*BLOCK_SIZE;
float *compressedBlock = blocks[I*NJ*NK + J*NK + K];
unsigned int csize = compressedSize[I*NJ*NK + J*NK + K];
float values[BLOCK_SIZE][BLOCK_SIZE][BLOCK_SIZE];
decompress(compressedBlock, csize, values);
return values[ii][jj][kk];
}
// number of blocks:
int NI, NJ, NK;
// number of samples:
int ni, nj, nk;
float** blocks;
unsigned int* compressedSize;
};
For this to be useful I need a lossy compression that is:
extremely fast, also on small datasets (e.g. 64x64x64)
compress quite hard > 3x, never mind if it looses quite a bit of info.
Any good candidates?
It sounds like you're looking for a LOD (level of detail) adaptive mesh. It's a recurring theme in video games and terrain simulation.
For terrain, see here: http://vterrain.org/LOD/Papers/ -- look for the ROAM video which is IIRC not only adaptive by distance, but also by view direction.
For non-terrain entities, there is a huge body of work (here's one example: Generic Adaptive Mesh Refinement).
I would suggest to use OctoMap to handle large 3D data.
And to extend it as shown here to handle geometrical properties.
I have to write in a PyCUDA function that gets two matrices Nx3 and Mx3, and return a matrix NxM, but I can't figure out how to pass by reference a matrix without knowing the number of columns.
My code basically is something like that:
#kernel declaration
mod = SourceModule("""
__global__ void distance(int N, int M, float d1[][3], float d2[][3], float res[][M])
{
int i = threadIdx.x;
int j = threadIdx.y;
float x, y, z;
x = d2[j][0]-d1[i][0];
y = d2[j][1]-d1[i][1];
z = d2[j][2]-d1[i][2];
res[i][j] = x*x + y*y + z*z;
}
""")
#load data
data1 = numpy.loadtxt("data1.txt").astype(numpy.float32) # Nx3 matrix
data2 = numpy.loadtxt("data2.txt").astype(numpy.float32) # Mx3 matrix
N=data1.shape[0]
M=data2.shape[0]
res = numpy.zeros([N,M]).astype(numpy.float32) # NxM matrix
#invoke kernel
dist_gpu = mod.get_function("distance")
dist_gpu(cuda.In(numpy.int32(N)), cuda.In(numpy.int32(M)), cuda.In(data1), cuda.In(data2), cuda.Out(res), block=(N,M,1))
#save data
numpy.savetxt("results.txt", res)
Compiling this I receive an error:
kernel.cu(3): error: a parameter is not allowed
that is, I cannot use M as the number of columns for res[][] in the declaretion of the function. I cannot either left the number of columns undeclared...
I need a matrix NxM as an output, but I can't figure out how to do this. Can you help me?
You should use pitched linear memory access inside the kernel, that is how ndarray and gpuarray store data internally, and PyCUDA will pass a pointer to the data in gpu memory allocated for a gpuarray when it is supplied as a argument to a PyCUDA kernel. So (if I understand what you are trying to do) your kernel should be written as something like:
__device__ unsigned int idx2d(int i, int j, int lda)
{
return j + i*lda;
}
__global__ void distance(int N, int M, float *d1, float *d2, float *res)
{
int i = threadIdx.x + blockDim.x * blockIdx.x;
int j = threadIdx.y + blockDim.y * blockIdx.y;
float x, y, z;
x = d2[idx2d(j,0,3)]-d1[idx2d(i,0,3)];
y = d2[idx2d(j,1,3)]-d1[idx2d(i,1,3)];
z = d2[idx2d(j,2,3)]-d1[idx2d(i,2,3)];
res[idx2d(i,j,N)] = x*x + y*y + z*z;
}
Here I have assumed the numpy default row major ordering in defining the idx2d helper function. There are still problems with the Python side of the code you posted, but I guess you know that already.
EDIT: Here is a complete working repro case based of the code posted in your question. Note that it only uses a single block (like the original), so be mindful of block and grid dimensions when trying to run it on anything other than trivially small cases.
import numpy as np
from pycuda import compiler, driver
from pycuda import autoinit
#kernel declaration
mod = compiler.SourceModule("""
__device__ unsigned int idx2d(int i, int j, int lda)
{
return j + i*lda;
}
__global__ void distance(int N, int M, float *d1, float *d2, float *res)
{
int i = threadIdx.x + blockDim.x * blockIdx.x;
int j = threadIdx.y + blockDim.y * blockIdx.y;
float x, y, z;
x = d2[idx2d(j,0,3)]-d1[idx2d(i,0,3)];
y = d2[idx2d(j,1,3)]-d1[idx2d(i,1,3)];
z = d2[idx2d(j,2,3)]-d1[idx2d(i,2,3)];
res[idx2d(i,j,N)] = x*x + y*y + z*z;
}
""")
#make data
data1 = np.random.uniform(size=18).astype(np.float32).reshape(-1,3)
data2 = np.random.uniform(size=12).astype(np.float32).reshape(-1,3)
N=data1.shape[0]
M=data2.shape[0]
res = np.zeros([N,M]).astype(np.float32) # NxM matrix
#invoke kernel
dist_gpu = mod.get_function("distance")
dist_gpu(np.int32(N), np.int32(M), driver.In(data1), driver.In(data2), \
driver.Out(res), block=(N,M,1), grid=(1,1))
print res
Does anybody know how to find the local maxima in a grayscale IPL_DEPTH_8U image using OpenCV? HarrisCorner mentions something like that but I'm actually not interested in corners ...
Thanks!
A pixel is considered a local maximum if it is equal to the maximum value in a 'local' neighborhood. The function below captures this property in two lines of code.
To deal with pixels on 'plateaus' (value equal to their neighborhood) one can use the local minimum property, since plateaus pixels are equal to their local minimum. The rest of the code filters out those pixels.
void non_maxima_suppression(const cv::Mat& image, cv::Mat& mask, bool remove_plateaus) {
// find pixels that are equal to the local neighborhood not maximum (including 'plateaus')
cv::dilate(image, mask, cv::Mat());
cv::compare(image, mask, mask, cv::CMP_GE);
// optionally filter out pixels that are equal to the local minimum ('plateaus')
if (remove_plateaus) {
cv::Mat non_plateau_mask;
cv::erode(image, non_plateau_mask, cv::Mat());
cv::compare(image, non_plateau_mask, non_plateau_mask, cv::CMP_GT);
cv::bitwise_and(mask, non_plateau_mask, mask);
}
}
Here's a simple trick. The idea is to dilate with a kernel that contains a hole in the center. After the dilate operation, each pixel is replaced with the maximum of it's neighbors (using a 5 by 5 neighborhood in this example), excluding the original pixel.
Mat1b kernelLM(Size(5, 5), 1u);
kernelLM.at<uchar>(2, 2) = 0u;
Mat imageLM;
dilate(image, imageLM, kernelLM);
Mat1b localMaxima = (image > imageLM);
Actually after I posted the code above I wrote a better and very very faster one ..
The code above suffers even for a 640x480 picture..
I optimized it and now it is very very fast even for 1600x1200 pic.
Here is the code :
void localMaxima(cv::Mat src,cv::Mat &dst,int squareSize)
{
if (squareSize==0)
{
dst = src.clone();
return;
}
Mat m0;
dst = src.clone();
Point maxLoc(0,0);
//1.Be sure to have at least 3x3 for at least looking at 1 pixel close neighbours
// Also the window must be <odd>x<odd>
SANITYCHECK(squareSize,3,1);
int sqrCenter = (squareSize-1)/2;
//2.Create the localWindow mask to get things done faster
// When we find a local maxima we will multiply the subwindow with this MASK
// So that we will not search for those 0 values again and again
Mat localWindowMask = Mat::zeros(Size(squareSize,squareSize),CV_8U);//boolean
localWindowMask.at<unsigned char>(sqrCenter,sqrCenter)=1;
//3.Find the threshold value to threshold the image
//this function here returns the peak of histogram of picture
//the picture is a thresholded picture it will have a lot of zero values in it
//so that the second boolean variable says :
// (boolean) ? "return peak even if it is at 0" : "return peak discarding 0"
int thrshld = maxUsedValInHistogramData(dst,false);
threshold(dst,m0,thrshld,1,THRESH_BINARY);
//4.Now delete all thresholded values from picture
dst = dst.mul(m0);
//put the src in the middle of the big array
for (int row=sqrCenter;row<dst.size().height-sqrCenter;row++)
for (int col=sqrCenter;col<dst.size().width-sqrCenter;col++)
{
//1.if the value is zero it can not be a local maxima
if (dst.at<unsigned char>(row,col)==0)
continue;
//2.the value at (row,col) is not 0 so it can be a local maxima point
m0 = dst.colRange(col-sqrCenter,col+sqrCenter+1).rowRange(row-sqrCenter,row+sqrCenter+1);
minMaxLoc(m0,NULL,NULL,NULL,&maxLoc);
//if the maximum location of this subWindow is at center
//it means we found the local maxima
//so we should delete the surrounding values which lies in the subWindow area
//hence we will not try to find if a point is at localMaxima when already found a neighbour was
if ((maxLoc.x==sqrCenter)&&(maxLoc.y==sqrCenter))
{
m0 = m0.mul(localWindowMask);
//we can skip the values that we already made 0 by the above function
col+=sqrCenter;
}
}
}
The following listing is a function similar to Matlab's "imregionalmax". It looks for at most nLocMax local maxima above threshold, where the found local maxima are at least minDistBtwLocMax pixels apart. It returns the actual number of local maxima found. Notice that it uses OpenCV's minMaxLoc to find global maxima. It is "opencv-self-contained" except for the (easy to implement) function vdist, which computes the (euclidian) distance between points (r,c) and (row,col).
input is one-channel CV_32F matrix, and locations is nLocMax (rows) by 2 (columns) CV_32S matrix.
int imregionalmax(Mat input, int nLocMax, float threshold, float minDistBtwLocMax, Mat locations)
{
Mat scratch = input.clone();
int nFoundLocMax = 0;
for (int i = 0; i < nLocMax; i++) {
Point location;
double maxVal;
minMaxLoc(scratch, NULL, &maxVal, NULL, &location);
if (maxVal > threshold) {
nFoundLocMax += 1;
int row = location.y;
int col = location.x;
locations.at<int>(i,0) = row;
locations.at<int>(i,1) = col;
int r0 = (row-minDistBtwLocMax > -1 ? row-minDistBtwLocMax : 0);
int r1 = (row+minDistBtwLocMax < scratch.rows ? row+minDistBtwLocMax : scratch.rows-1);
int c0 = (col-minDistBtwLocMax > -1 ? col-minDistBtwLocMax : 0);
int c1 = (col+minDistBtwLocMax < scratch.cols ? col+minDistBtwLocMax : scratch.cols-1);
for (int r = r0; r <= r1; r++) {
for (int c = c0; c <= c1; c++) {
if (vdist(Point2DMake(r, c),Point2DMake(row, col)) <= minDistBtwLocMax) {
scratch.at<float>(r,c) = 0.0;
}
}
}
} else {
break;
}
}
return nFoundLocMax;
}
The first question to answer would be what is "local" in your opinion. The answer may well be a square window (say 3x3 or 5x5) or circular window of a certain radius. You can then scan over the entire image with the window centered at each pixel and pick the highest value in the window.
See this for how to access pixel values in OpenCV.
This is very fast method. It stored founded maxima in a vector of
Points.
vector <Point> GetLocalMaxima(const cv::Mat Src,int MatchingSize, int Threshold, int GaussKernel )
{
vector <Point> vMaxLoc(0);
if ((MatchingSize % 2 == 0) || (GaussKernel % 2 == 0)) // MatchingSize and GaussKernel have to be "odd" and > 0
{
return vMaxLoc;
}
vMaxLoc.reserve(100); // Reserve place for fast access
Mat ProcessImg = Src.clone();
int W = Src.cols;
int H = Src.rows;
int SearchWidth = W - MatchingSize;
int SearchHeight = H - MatchingSize;
int MatchingSquareCenter = MatchingSize/2;
if(GaussKernel > 1) // If You need a smoothing
{
GaussianBlur(ProcessImg,ProcessImg,Size(GaussKernel,GaussKernel),0,0,4);
}
uchar* pProcess = (uchar *) ProcessImg.data; // The pointer to image Data
int Shift = MatchingSquareCenter * ( W + 1);
int k = 0;
for(int y=0; y < SearchHeight; ++y)
{
int m = k + Shift;
for(int x=0;x < SearchWidth ; ++x)
{
if (pProcess[m++] >= Threshold)
{
Point LocMax;
Mat mROI(ProcessImg, Rect(x,y,MatchingSize,MatchingSize));
minMaxLoc(mROI,NULL,NULL,NULL,&LocMax);
if (LocMax.x == MatchingSquareCenter && LocMax.y == MatchingSquareCenter)
{
vMaxLoc.push_back(Point( x+LocMax.x,y + LocMax.y ));
// imshow("W1",mROI);cvWaitKey(0); //For gebug
}
}
}
k += W;
}
return vMaxLoc;
}
Found a simple solution.
In this example, if you are trying to find 2 results of a matchTemplate function with a minimum distance from each other.
cv::Mat result;
matchTemplate(search, target, result, CV_TM_SQDIFF_NORMED);
float score1;
cv::Point displacement1 = MinMax(result, score1);
cv::circle(result, cv::Point(displacement1.x+result.cols/2 , displacement1.y+result.rows/2), 10, cv::Scalar(0), CV_FILLED, 8, 0);
float score2;
cv::Point displacement2 = MinMax(result, score2);
where
cv::Point MinMax(cv::Mat &result, float &score)
{
double minVal, maxVal;
cv::Point minLoc, maxLoc, matchLoc;
minMaxLoc(result, &minVal, &maxVal, &minLoc, &maxLoc, cv::Mat());
matchLoc.x = minLoc.x - result.cols/2;
matchLoc.y = minLoc.y - result.rows/2;
return minVal;
}
The process is:
Find global Minimum using minMaxLoc
Draw a filled white circle around global minimum using min distance between minima as radius
Find another minimum
The the scores can be compared to each other to determine, for example, the certainty of the match,
To find more than just the global minimum and maximum try using this function from skimage:
http://scikit-image.org/docs/dev/api/skimage.feature.html#skimage.feature.peak_local_max
You can parameterize the minimum distance between peaks, too. And more. To find minima, use negated values (take care of the array type though, 255-image could do the trick).
You can go over each pixel and test if it is a local maxima. Here is how I would do it.
The input is assumed to be type CV_32FC1
#include <vector>//std::vector
#include <algorithm>//std::sort
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/core/core.hpp"
//structure for maximal values including position
struct SRegionalMaxPoint
{
SRegionalMaxPoint():
values(-FLT_MAX),
row(-1),
col(-1)
{}
float values;
int row;
int col;
//ascending order
bool operator()(const SRegionalMaxPoint& a, const SRegionalMaxPoint& b)
{
return a.values < b.values;
}
};
//checks if pixel is local max
bool isRegionalMax(const float* im_ptr, const int& cols )
{
float center = *im_ptr;
bool is_regional_max = true;
im_ptr -= (cols + 1);
for (int ii = 0; ii < 3; ++ii, im_ptr+= (cols-3))
{
for (int jj = 0; jj < 3; ++jj, im_ptr++)
{
if (ii != 1 || jj != 1)
{
is_regional_max &= (center > *im_ptr);
}
}
}
return is_regional_max;
}
void imregionalmax(
const cv::Mat& input,
std::vector<SRegionalMaxPoint>& buffer)
{
//find local max - top maxima
static const int margin = 1;
const int rows = input.rows;
const int cols = input.cols;
for (int i = margin; i < rows - margin; ++i)
{
const float* im_ptr = input.ptr<float>(i, margin);
for (int j = margin; j < cols - margin; ++j, im_ptr++)
{
//Check if pixel is local maximum
if ( isRegionalMax(im_ptr, cols ) )
{
cv::Rect roi = cv::Rect(j - margin, i - margin, 3, 3);
cv::Mat subMat = input(roi);
float val = *im_ptr;
//replace smallest value in buffer
if ( val > buffer[0].values )
{
buffer[0].values = val;
buffer[0].row = i;
buffer[0].col = j;
std::sort(buffer.begin(), buffer.end(), SRegionalMaxPoint());
}
}
}
}
}
For testing the code you can try this:
cv::Mat temp = cv::Mat::zeros(15, 15, CV_32FC1);
temp.at<float>(7, 7) = 1;
temp.at<float>(3, 5) = 6;
temp.at<float>(8, 10) = 4;
temp.at<float>(11, 13) = 7;
temp.at<float>(10, 3) = 8;
temp.at<float>(7, 13) = 3;
vector<SRegionalMaxPoint> buffer_(5);
imregionalmax(temp, buffer_);
cv::Mat debug;
cv::cvtColor(temp, debug, cv::COLOR_GRAY2BGR);
for (auto it = buffer_.begin(); it != buffer_.end(); ++it)
{
circle(debug, cv::Point(it->col, it->row), 1, cv::Scalar(0, 255, 0));
}
This solution does not take plateaus into account so it is not exactly the same as matlab's imregionalmax()
I think you want to use the
MinMaxLoc(arr, mask=NULL)-> (minVal, maxVal, minLoc, maxLoc)
Finds global minimum and maximum in array or subarray
function on you image