joaofaro / kcfcpp Goto Github PK

Can anyone tell me if this method is faster than opencv KCF? How faster?

hi,
compile error: bool KCFTracker::createHanningMats() in kcftracker.cpp have no return value .

I test the code on some videos captured by myself , then I found that the tracking results varies greatly when I just initialize the algorithm with a slightly larger or smaller target box.
I wonder how can this happen since the code always extract a larger roi and get the feature, and I thought it will not so sensitive to the size of the box

I use the KCV to track the moving object.When I set HOG false,it behaves very well,and peak_value set originally is 2.When set HOG false,the compiler returns the following errors.

OpenCV Error: Assertion failed (src.size() == dst.size() && src.channels() == dst.channels()) in cvFilter2D, file /opencv-2.4.13/opencv/modules/imgproc/src/filter.cpp, line 3359
terminate called after throwing an instance of 'cv::Exception'
what(): /opencv-2.4.13/opencv/modules/imgproc/src/filter.cpp:3359: error: (-215) src.size() == dst.size() && src.channels() == dst.channels() in function cvFilter2D.

I check the cvFilter2D,and image I use to init the KCFTracker and used to update the Rect are all gray images, and the size are also right.I am curious.
Do some friends know the reasons?
Thank you in advance!

I need to judge whether the target is lost after the prediction, how to judge whether the target is lost or not？

Hello,

I always have this problem when building the project in debug mode.
It looks like a memory error comes from the real part of the fft().

cv::Mat real(cv::Mat img)
{
    std::vector<cv::Mat> planes;
    cv::split(img, planes);                     // Error occurs here.
    return planes[0];
}

This problem disappear in release mode. The compiler is VS2013.

When a object go away and it become smaller，but the finding box still big？how to overcome it?
chinese : 当追踪目标远离时，发现框的矩形大小不会变化？为什么？怎么修改？thanks😊😉

Thanks a lot for your open source, you are really a kind man. Recently, I have read your two papers, "Exploiting the Circulant Structure of Tracking-by-detection with Kernels" and "High-Speed Tracking with Kernelized Correlation Filters", and I download your codes (Matlab version) from the hyperlink given by the "Visual Tracker Benchmark", it's really very useful for me, a novice, to get further understanding about the KCF algorithm. Unfortunately, there are still some questions for me to understanding the algorithm and the codes, that I don't understand the dense sampling and the usefulness of the command 'circshift' in your "Code Version 1", the "dense_gauss_kernel.m" more specifically,
xy = real(circshift(ifft2(xyf), floor(size(x)/2))); %to spatial domain
http://www.robots.ox.ac.uk/~joao/circulant/index.html(I get your code from this url, Matlab code — version 1)

So I wish to get some instructions from you to help me understand this problem better.I would appreciate it very much if i can obtain any support from you.

when i set a target size, the roipadding is the _tmp's size, because of the hog size is more less than the source image, the hog size should be (roipadding)/cell_size, but in this project i find that the hog size is more less that the size i think when i set the target a large size(for example: the target size is 112115,padding=2.5 the hog size is just 2224?? ), what is the problem?

** linux, without VOT toolkit , I comment out some lines of code in runtracker.cpp as following,**
`#include
#include
#include
#include

#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>

#include "kcftracker.hpp"

#include <dirent.h>

using namespace std;
using namespace cv;

int main(int argc, char* argv[]){

if (argc > 5) return -1;

bool HOG = true;
bool FIXEDWINDOW = false;
bool MULTISCALE = true;
bool SILENT = true;
bool LAB = false;

for(int i = 0; i < argc; i++){
	if ( strcmp (argv[i], "hog") == 0 )
		HOG = true;
	if ( strcmp (argv[i], "fixed_window") == 0 )
		FIXEDWINDOW = true;
	if ( strcmp (argv[i], "singlescale") == 0 )
		MULTISCALE = false;
	if ( strcmp (argv[i], "show") == 0 )
		SILENT = false;
	if ( strcmp (argv[i], "lab") == 0 ){
		LAB = true;
		HOG = true;
	}
	if ( strcmp (argv[i], "gray") == 0 )
		HOG = false;
}

// Create KCFTracker object
KCFTracker tracker(HOG, FIXEDWINDOW, MULTISCALE, LAB);

// Frame readed
Mat frame;

// Tracker results
Rect result;

/* // Path to list.txt
ifstream listFile;
string fileName = "images.txt";
listFile.open(fileName);

// Read groundtruth for the 1st frame
ifstream groundtruthFile;
string groundtruth = "region.txt";
groundtruthFile.open(groundtruth);
string firstLine;
getline(groundtruthFile, firstLine);
groundtruthFile.close();

istringstream ss(firstLine);

*/

// Read groundtruth like a dumb
float x1, y1, x2, y2, x3, y3, x4, y4;
char ch;
/*ss >> x1;
ss >> ch;
ss >> y1;
ss >> ch;
ss >> x2;
ss >> ch;
ss >> y2;
ss >> ch;
ss >> x3;
ss >> ch;
ss >> y3;
ss >> ch;
ss >> x4;
ss >> ch;
ss >> y4;      */

// Using min and max of X and Y for groundtruth rectangle
float xMin =  min(x1, min(x2, min(x3, x4)));
float yMin =  min(y1, min(y2, min(y3, y4)));
float width = max(x1, max(x2, max(x3, x4))) - xMin;
float height = max(y1, max(y2, max(y3, y4))) - yMin;


// Read Images
ifstream listFramesFile;
/*string listFrames = "images.txt";
listFramesFile.open(listFrames);    */
string frameName;


// Write Results
ofstream resultsFile;
string resultsPath = "output.txt";
resultsFile.open(resultsPath);

// Frame counter
int nFrames = 0;


while ( getline(listFramesFile, frameName) ){
	frameName = frameName;

	// Read each frame from the list
	frame = imread(frameName, CV_LOAD_IMAGE_COLOR);

	// First frame, give the groundtruth to the tracker
	if (nFrames == 0) {
		tracker.init( Rect(xMin, yMin, width, height), frame );
		rectangle( frame, Point( xMin, yMin ), Point( xMin+width, yMin+height), Scalar( 0, 255, 255 ), 1, 8 );
		resultsFile << xMin << "," << yMin << "," << width << "," << height << endl;
	}
	// Update
	else{
		result = tracker.update(frame);
		rectangle( frame, Point( result.x, result.y ), Point( result.x+result.width, result.y+result.height), Scalar( 0, 255, 255 ), 1, 8 );
		resultsFile << result.x << "," << result.y << "," << result.width << "," << result.height << endl;
	}

	nFrames++;

	if (!SILENT){
		imshow("Image", frame);
		waitKey(1);
	}
}
resultsFile.close();

// listFile.close();

}
**# What should I do to show the result correctly? thank u.**

What the function update returns if object disappears, say occluded for a while? Is there any way to return confidence?

1>f:\zhweizh\image processing\vs prj\kcf\kcf\src\runtracker.cpp(11): fatal error C1083: 无法打开包括文件:“dirent.h”: No such file or directory. in vs2010+opencv244 enviroment compiling
look like there is no dirent.h file.

Looking at this part (update method), it seems the _roi.width, _roi.height, and _scale will not change to bigger scale when bigger_scale_peak > smaller_scale_peak > normal_scale_peak. (I removed the scale_weight multiplication in this explanation to keep it simple)

Because the variables are first divided by scale_step here then multiplied with the same scale_step here. In the end, the variables return to normal scale in this case, while it should be bigger scale.

CMIIW. Thank you~

Can we use KCF to Multiple Object Tracking?

When running in Docker container, failed with

libdc1394 error: Failed to initialize libdc1394

My opencv is opencv4.2, and I get this problem:
IplImage z_ipl = z;
conversion from ‘cv::Mat’ to non-scalar type ‘IplImage {aka _IplImage}’ requested

How can I solve this problem. Thanks a lot.

Hello@joaofaro:
can you tell me whether this code can be used in the windows,Thank you!

great project!!! i want to try the fftw support option,but when i turn it on with "-DUSE_FFTW", i got a runtime failed:

any idea ??

Hi,

When I create a KCFTracker object like this
`KCFTracker tracker(HOG, FIXEDWINDOW, MULTISCALE, LAB), I'm getting the following error when building the project. I'm using VS2012. Could anyone please help ? Thanks in advance.

error LNK2001: unresolved external symbol "public: __thiscall KCFTracker::KCFTracker(bool,bool,bool,bool)" (??0KCFTracker@@QAE@_N000@Z)

When I checked detect method's both double pv and float peak_value variables after peak_value = (float) pv; line, sometimes I got value more than 1.0 (mostly still less than 1.1 but can be higher). Have any idea what would be the problem?

Edit:
it happens when HOG is set False

I find only Gaussian kernels. According to "High-Speed Tracking with Kernelized Correlation Filters" , linear kernel(DCF) is much faster than KCF (292FPS vs 172FPS).

to do a complex division, say
(a+bi)/(c+di) = (a+bi)(c-di)/(cc+dd)
= [(ac+bd)+(bc-ad)i]/(cc+dd)
so in the code

    pres.push_back((pa[0].mul(pb[0]) + pa[1].mul(pb[1])).mul(divisor));
    pres.push_back((pa[1].mul(pb[0]) + pa[0].mul(pb[1])).mul(divisor));

second '+' should be '-'

But why does the code still work?