Ask Your Question
0

Hog detector for hand recognition

asked 2017-11-15 13:08:47 -0500

xavier12358 gravatar image

updated 2017-11-17 02:17:51 -0500

Hello,

I am trying to detect hand in images thanks to Hog detection and SVM network.

Is it a good idea? Because I try with dataset and it doesn't work properly ...

Edit:

So basically I use that code: https://github.com/lcit/people_detect...

I just modify the database I use , I used that one: https://expirebox.com/download/f61af2...

I modify the training cpp file like this:

/*  =========================================================================
    Author: Leonardo Citraro
    Company:
    Filename: training.cpp
    Last modifed:   22.12.2016 by Leonardo Citraro
    Description:    Training of the classifier using the HOG feature

    =========================================================================

    =========================================================================
*/
#include "HOG.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/ml/ml.hpp"
#include "opencv2/core/utility.hpp"
#include "opencv2/imgcodecs.hpp"
#include <iostream>
#include <fstream>
#include <vector>
#include <algorithm>
#include <memory>
#include <random>
#include <functional>
#include <ctime>
#include <iomanip>
#include <math.h>

static int MatTYPE = CV_32FC1;
using TYPE = float;

TYPE compute_mean(std::vector<TYPE> v) {
    return std::accumulate(std::begin(v), std::end(v), 0.0f)/v.size();
}

void feature_mean_variance(const cv::Mat& data, std::vector<TYPE>& mean, std::vector<TYPE>& var) {
    mean.resize(data.cols);
    var.resize(data.cols);

    for(size_t col=0; col<data.cols; ++col) {
        std::vector<TYPE> feature(data.rows);
        for(size_t i = 0; i < data.rows; ++i) {
            const TYPE* ptr_row = data.ptr<TYPE>(i);
            feature[i] = ptr_row[col];
        }
        TYPE m = std::accumulate(std::begin(feature), std::end(feature), 0.0)/feature.size();
        mean[col] = m;
        std::vector<TYPE> diff(data.rows);
        std::transform(std::begin(feature), std::end(feature), std::begin(diff), std::bind2nd(std::minus<TYPE>(), m));
        TYPE v = std::inner_product(std::begin(diff), std::end(diff), std::begin(diff), 0.0)/feature.size();
        var[col] = v;
    }
}

template<class T>
void save_vector( const std::string& filename, const std::vector<T>& v ) {
    try {
        std::ofstream f(filename, std::ios::binary);
        unsigned int len = v.size();
        f.write( (char*)&len, sizeof(len) );
        f.write( (const char*)&v[0], len * sizeof(T) );
        f.close();
    } catch(...) {
        throw;
    }
}

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

    // size of the box that should contain a person
    cv::Size person_size(50,75);

    // setting up the HOG
    size_t cellsize = 5;
    size_t blocksize = cellsize*2;
    size_t stride = cellsize;
    size_t binning = 9;
    HOG hog(blocksize, cellsize, stride, binning, HOG::GRADIENT_SIGNED, HOG::BLOCK_NORM::L2norm);
    hog.save("hog.ext");

    // matrix of data and labels
    std::vector<std::vector<TYPE>> data;
    std::vector<int> labels;

    // open the subimages of the persons one by one
    for(int i=1; i<400; ++i){
    std::string filename;
    if(i<10){
        filename = "/home/xavier/Bureau/developpement/NeuralNetwork/dataset/hand/Marcel-Train/Five/Five-train00" + std::to_string(i) + ".ppm";
    }else if(i<100){
        filename = "/home/xavier/Bureau/developpement/NeuralNetwork/dataset/hand/Marcel-Train/Five/Five-train0" + std::to_string(i) + ".ppm";
    }else if(i<1000){
        filename = "/home/xavier/Bureau/developpement/NeuralNetwork/dataset/hand/Marcel-Train/Five/Five-train" + std::to_string(i) + ".ppm";
    }


        try {
        //std::cout << "debut filename" << filename << std::endl;
            // open and display an image
            cv::Mat image = cv::imread(filename, CV_8U);
        cv::resize(image,image,person_size);
        /*cv::imshow("image",image);
        cv::waitKey(-1);*/
            if(image.data) {
        //std::cout << "la0" << std::endl;
                // Retrieve the ...
(more)
edit retag flag offensive close merge delete

Comments

How about this one? hog detector

supra56 gravatar imagesupra56 ( 2017-11-15 19:50:42 -0500 )edit

Consider showing exactly what you have tried and your results thus far. It will be hard for others to assist without you providing more information. Additionally, if you show effort in describing what you are trying to solve, others will be far more inclined to assist you.

Der Luftmensch gravatar imageDer Luftmensch ( 2017-11-16 08:33:13 -0500 )edit

you should probably use opencv's training tool for this.

above code looks quite shoddy. (code repetition, it's also doing a classification, while you need a regression for the detection case)

and, my 2ct.: unless you restrict it to a single pose, chances of success are low.

berak gravatar imageberak ( 2017-11-17 02:40:26 -0500 )edit

The images should have a multiply of 8 for the size?

xavier12358 gravatar imagexavier12358 ( 2017-11-17 04:04:48 -0500 )edit

not sure, but probably yes.

berak gravatar imageberak ( 2017-11-17 04:08:10 -0500 )edit

I have also wrong behaviors. What should be the size of the dataset?

xavier12358 gravatar imagexavier12358 ( 2017-11-17 04:34:50 -0500 )edit

you mean, the size of the images ? the positives should be cropped /resized to the HOGDetector's winSize.

(that's also the minimum size, that can be detected later)

the negatives might be large, it will automatically sample a region from that

berak gravatar imageberak ( 2017-11-17 05:04:13 -0500 )edit

Where could I get example of use of that train_HOG example?

xavier12358 gravatar imagexavier12358 ( 2017-11-17 06:11:44 -0500 )edit

what do you mean ? a pretrained svm detector ?

berak gravatar imageberak ( 2017-11-17 06:51:21 -0500 )edit
1

maybe you can find this post and this repo useful

sturkmen gravatar imagesturkmen ( 2017-11-17 08:03:18 -0500 )edit

1 answer

Sort by ยป oldest newest most voted
1

answered 2017-11-18 15:11:56 -0500

i trained a hand detector naively using some files of your dataset. i just aimed to show you how to select positive images (you can find them in here).

you can test my hand detector or train yours using train_HOG.cpp ( i recently updated it according to your dataset and added ability of testing with webcam )

image description

edit flag offensive delete link more

Comments

What are the negative example you selected? Did you try that with more complexe background ?

xavier12358 gravatar imagexavier12358 ( 2017-11-18 22:02:07 -0500 )edit

i used INRIAPerson\neg folder as negative image folder. (it contains 1148 images) you can use any image not containing a hand as negative.

when the background is complex. detection is very poor.

sturkmen gravatar imagesturkmen ( 2017-11-19 10:48:24 -0500 )edit
Login/Signup to Answer

Question Tools

2 followers

Stats

Asked: 2017-11-15 13:08:47 -0500

Seen: 559 times

Last updated: Nov 18 '17