Latent SVM is a detector that uses HOG features and a star-structured, part-based model consisting of a root filter and a set of part filters to represent an object category. HOGs are feature descriptors that are obtained by counting the occurrences of gradient orientations in localized portions of an image. On the other hand, a variant of support vector machines (SVM) classifiers are used in this detector to train models using partially labeled data. The basic idea of an SVM is constructing a hyperplane or set of hyperplanes in high-dimensional space. These hyperplanes are obtained to have the largest distance to the nearest training data point (functional margin in order to achieve low generalization errors). Like cascade detectors, Latent SVM uses a sliding window with different initial positions and scales where the algorithm is applied in order to detect if there is an object inside.

One of the advantages of the OpenCV Latent SVM implementation is that it allows the detection of multiple object categories by combining several simple pretrained detectors within the same multiobject detector instance.

The following latentDetection example illustrates how to use a Latent SVM detector for localizing objects from a category in an image:

#include "opencv2/core/core.hpp"
#include "opencv2/objdetect/objdetect.hpp"
#include "opencv2/highgui/highgui.hpp"
#include <iostream>

using namespace std;
using namespace cv;

int main(int argc, char* argv[]){
    String model = argv[1];
    vector<String> models;
    models.push_back( model );
    vector<String> names;
    names.push_back( "category" );
    LatentSvmDetector detector( models , names);
    if( detector.empty() ) {
        cout << "Model cannot be loaded" << endl;
        return -1;
    }

    String img = argv[2];
    Mat image = imread( img );
    if( image.empty() ){
        cout << "Image cannot be loaded" << endl;
        return -1;
    }

    vector<LatentSvmDetector::ObjectDetection> detections;
    detector.detect( image, detections, 0.1, 1);
    for( size_t i = 0; i < detections.size(); i++ ) {
        Point center( detections[i].rect.x + 
                      detections[i].rect.width*0.5, 
                      detections[i].rect.y + 
                      detections[i].rect.height*0.5 );
        ellipse( image, center, Size( detections[i].rect.width*0.5, 
                 detections[i].rect.height*0.5), 0, 0, 360, 
                 Scalar( 255, 0, 255 ), 4, 8, 0 );
    }
    imshow( "result", image );
    waitKey(0);
    return 0;
}

The code explanation is as follows:

The following screenshot shows a cat detected using the previous code and the files cat.xml and cat.png, and cars detected using car.xml and cars.png. These files are included in the OpenCV extra data that can be found in the official repository. Thus, it is possible to run the program using the following command:

>latentDetection.exe xmlfile imagefile

In the previous command, xmlfile is the Latent SVM detector and imagefile is the image that has to be examined.

In addition to the car and cat detectors, OpenCV provides pretrained detectors for the rest of the classes defined in The PASCAL Visual Object Classes Challenge 2007 (http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2007). These detectors are as follows:

The detection of a cat and some cars using Latent SVM