OpenCV Latent SVM Discriminatively Trained Part Based Models for Object Detection

原文：http://blog.csdn.net/loadstar_kun/article/details/8686416

OpenCV 2.4版本实现了C++版本的DPM程序，和之前C版本的主要区别就是可以同时检测多种目标。使用时，可以把训练好的模型放到一个文件夹中，待检测的图像放到另一个文件夹中，即可检测。

可惜没有考虑加速的内容。

 

Latent SVM¶

Discriminatively Trained Part Based Models for Object Detection¶

The object detector described below has been initially proposed by P.F. Felzenszwalb in[Felzenszwalb2010]. 
It is based on a Dalal-Triggs detector that uses a single filter on histogram of oriented gradients (HOG) features to represent an object category. This detector uses a sliding window approach, where a filter is applied at all positions and scales of an image.
The first innovation is enriching the Dalal-Triggs model using a star-structured part-based model defined by a “root” filter (analogous to the Dalal-Triggs filter) plus a set of parts filters and associated deformation models. The score of one of star models
at a particular position and scale within an image is the score of the root filter at the given location plus the sum over parts of the maximum, over placements of that part, of the part filter score on its location minus a deformation cost easuring the deviation
of the part from its ideal location relative to the root. Both root and part filter scores are defined by the dot product between a filter (a set of weights) and a subwindow of a feature pyramid computed from the input image. Another improvement is a representation
of the class of models by a mixture of star models. The score of a mixture model at a particular position and scale is the maximum over components, of the score of that component model at the given location.

In OpenCV there are C implementation of Latent SVM and C++ wrapper of it. C version is the structureCvObjectDetection and
a set of functions working with this structure (seecvLoadLatentSvmDetector(),cvReleaseLatentSvmDetector(),cvLatentSvmDetectObjects()).
C++ version is the classLatentSvmDetector and
has slightly different functionality in contrast with C version - it supports loading and detection of several models.

There are two examples of Latent SVM usage:samples/c/latentsvmdetect.cppandsamples/cpp/latentsvm_multidetect.cpp.

 

代码如下：

[html] view
plaincopy

#include <iostream>  

#include "opencv2/objdetect/objdetect.hpp"  

#include "opencv2/highgui/highgui.hpp"  

#include "opencv2/contrib/contrib.hpp"  

  

#ifdef WIN32  

#include <io.h>  

#else  

#include <dirent.h>  

#endif  

  

#ifdef HAVE_CVCONFIG_H  

#include <cvconfig.h>  

#endif  

  

#ifdef HAVE_TBB  

#include "tbb/task_scheduler_init.h"  

#endif  

  

using namespace std;  

using namespace cv;  

  

static void help()  

{  

    cout << "This program demonstrated the use of the latentSVM detector." << endl <<  

            "It reads in a trained object models and then uses them to detect the objects in an images." << endl <<  

             endl <<  

            "Call:" << endl <<  

            "./latentsvm_multidetect <imagesFolder> <modelsFolder> [<overlapThreshold>][<threadsNumber>]" << endl <<  

            "<overlapThreshold> - threshold for the non-maximum suppression algorithm." << endl <<  

            "Example of <modelsFolder> is opencv_extra/testdata/cv/latentsvmdetector/models_VOC2007" << endl <<  

             endl <<  

            "Keys:" << endl <<  

            "'n' - to go to the next image;" << endl <<  

            "'esc' - to quit." << endl <<  

            endl;  

}  

  

static void detectAndDrawObjects( Mat& image, LatentSvmDetector& detector, const vector<Scalar>& colors, float overlapThreshold, int numThreads )  

{  

    vector<LatentSvmDetector::ObjectDetection> detections;  

  

    TickMeter tm;  

    tm.start();  

    detector.detect( image, detections, overlapThreshold, numThreads);  

    tm.stop();  

  

    cout << "Detection time = " << tm.getTimeSec() << " sec" << endl;  

  

    const vector<string> classNames = detector.getClassNames();  

    CV_Assert( colors.size() == classNames.size() );  

  

    for( size_t i = 0; i < detections.size(); i++ )  

    {  

        const LatentSvmDetector::ObjectDetection& od = detections[i];  

        rectangle( image, od.rect, colors[od.classID], 3 );  

    }  

    // put text over the all rectangles  

    for( size_t i = 0; i < detections.size(); i++ )  

    {  

        const LatentSvmDetector::ObjectDetection& od = detections[i];  

        putText( image, classNames[od.classID], Point(od.rect.x+4,od.rect.y+13), FONT_HERSHEY_SIMPLEX, 0.55, colors[od.classID], 2 );  

    }  

}  

  

static void readDirectory( const string& directoryName, vector<string>& filenames, bool addDirectoryName=true )  

{  

    filenames.clear();  

  

#ifdef WIN32  

    struct _finddata_t s_file;  

    string str = directoryName + "\\*.*";  

  

    intptr_t h_file = _findfirst( str.c_str(), &s_file );  

    if( h_file != static_cast<intptr_t>(-1.0) )  

    {  

        do  

        {  

            if( addDirectoryName )  

                filenames.push_back(directoryName + "\\" + s_file.name);  

            else  

                filenames.push_back((string)s_file.name);  

        }  

        while( _findnext( h_file, &s_file ) == 0 );  

    }  

    _findclose( h_file );  

#else  

    DIR* dir = opendir( directoryName.c_str() );  

    if( dir != NULL )  

    {  

        struct dirent* dent;  

        while( (dent = readdir(dir)) != NULL )  

        {  

            if( addDirectoryName )  

                filenames.push_back( directoryName + "/" + string(dent->d_name) );  

            else  

                filenames.push_back( string(dent->d_name) );  

        }  

    }  

#endif  

  

    sort( filenames.begin(), filenames.end() );  

}  

  

int main(int argc, char* argv[])  

{  

    help();  

  

    string images_folder, models_folder;  

    float overlapThreshold = 0.2f;  

    int numThreads = -1;  

    if( argc > 2 )  

    {  

        images_folder = argv[1];  

        models_folder = argv[2];  

        if( argc > 3 ) overlapThreshold = (float)atof(argv[3]);  

        if( overlapThreshold < 0 || overlapThreshold > 1)  

        {  

            cout << "overlapThreshold must be in interval (0,1)." << endl;  

            exit(-1);  

        }  

  

        if( argc > 4 ) numThreads = atoi(argv[4]);  

    }  

  

    vector<string> images_filenames, models_filenames;  

    readDirectory( images_folder, images_filenames );  

    readDirectory( models_folder, models_filenames );  

  

    LatentSvmDetector detector( models_filenames );  

    if( detector.empty() )  

    {  

        cout << "Models cann't be loaded" << endl;  

        exit(-1);  

    }  

  

    const vector<string>& classNames = detector.getClassNames();  

    cout << "Loaded " << classNames.size() << " models:" << endl;  

    for( size_t i = 0; i < classNames.size(); i++ )  

    {  

        cout << i << ") " << classNames[i] << "; ";  

    }  

    cout << endl;  

  

    cout << "overlapThreshold = " << overlapThreshold << endl;  

  

    vector<Scalar> colors;  

    generateColors( colors, detector.getClassNames().size() );  

  

    for( size_t i = 0; i < images_filenames.size(); i++ )  

    {  

        Mat image = imread( images_filenames[i] );  

        if( image.empty() )  continue;  

  

        cout << "Process image " << images_filenames[i] << endl;  

        detectAndDrawObjects( image, detector, colors, overlapThreshold, numThreads );  

  

        imshow( "result", image );  

  

        for(;;)  

        {  

            int c = waitKey();  

            if( (char)c == 'n')  

                break;  

            else if( (char)c == '\x1b' )  

                exit(0);  

        }  

    }  

  

    return 0;  

}  

 上传个结果吧。