blink.cpp

#include <opencv2/opencv.hpp>
#include <iostream>
#include <deque>
#include <map>
using namespace cv;
using namespace std;


int main(int argc, char **argv)
{
    string blink = "C:/data/blink/";
    string dummy;

    // there are 5000 frames in the db, first assume
    // they're all "non-blink":
    vector<int> tlabels(15832,0);
    // then fill in the ground-truth positives:
    ifstream anno(blink + "blink8_groundtruth.txt");
    getline(anno, dummy);
	

			 
    while ( getline (anno,dummy) )
		{
        int a;
        anno >> a;
        anno >> dummy ;
        if (dummy == "close")
         { // only "Full blink used here. data file contains frame, state (close)"
            fill(tlabels.begin() + a, 1);
        }
			// for negative blink (half).
			else
			{
				fill(tlabels.begin() + a, 0);
			}
    }

    // parse 5000 annotation(landmarks) files,
    // select p27,p28,p29,p30 for left eye,
    // and    p32,p33,p34,p35 for right eye.
    // calculate eye-aspect-ratio (ear) for both,
    // and write a 13-element time-series to csv.
    ofstream csv("ear.csv");
    deque<float> el,er;
	
    for (size_t i=0; i<15832; i++) 
	{
        ifstream in_L(blink + "EAR_OUTPUT_left_EYE.txt");
        if (! in_L.good()) {
            cerr << "bad file " << i << endl;
            continue;
        }
        // skip first 30 lines (3 lines offset + 27) , why it is 30 ? is this the case for blink8 data ?
        string dummy_L;
        for (int j=0; j<15832; j++)
	{ 
			
            getline(in_L, dummy_L);
           el.push_back(ear_l);
		
	}

		
   for (size_t i=0; i<15832; i++) 
	{
        ifstream in_R(blink + "EAR_OUTPUT_RIGHT_EYE.txt");
        if (! in_R.good()) {
            cerr << "bad file " << i << endl;
            continue;
        }
        // skip first 30 lines (3 lines offset + 27) , why it is 30 ? is this the case for blink8 data ?
        string dummy_R;
        for (int j=0; j<15832; j++)
	{ 
			
            getline(in_R, dummy_R);
           el.push_back(ear_r);
		
	}

        

        // write out queues for left & right eye
        // "backdated" to t-6 ! (center of the 13 elem queue)
        if (i >= 13) {
            csv << tlabels[i-6] << ",";
            for (int k=0; k<13; k++)
                csv << el[k] << ",";
            csv << endl;

            csv << tlabels[i-6] << ",";
            for (int k=0; k<13; k++)
                csv << er[k] << ",";
            csv << endl;

            el.pop_front();
            er.pop_front();
        }
    }

    // train & test with SVM:
    Ptr<ml::SVM> svm = ml::SVM::create();
    Ptr<ml::TrainData> tdata = ml::TrainData::loadFromCSV("ear.csv",0,0,1);
    tdata->setTrainTestSplitRatio(0.7);

    Mat data = tdata->getTrainSamples();
    Mat labels = tdata->getTrainResponses();
    labels.convertTo(labels, CV_32S); // hrrm!
    svm->train(data,0,labels);

    Mat vdata = tdata->getTestSamples();
    Mat vlabels = tdata->getTestResponses();
    Mat results;
    svm->predict(vdata,results);

    float correct = countNonZero(results == vlabels);
    float accuracy = correct / results.total();
    cerr << "accuracy: " << accuracy << endl;

    // accuracy alone is not enough here, since it might
    // simply have missed all positives !
    Mat_<float> confusion(2,2);
    for (int i=0; i<results.rows; i++) {
        int p = (int)results.at<float>(i);
        int t = (int)vlabels.at<float>(i);
        confusion(p,t) ++;
    }
    cerr << "confusion:\n" << confusion << endl;

    // additionally, do ROC analysis.
    // we need raw output, so another prediction required:
    svm->predict(vdata, results, ml::StatModel::RAW_OUTPUT);
    normalize(results, results, 1, 0, NORM_MINMAX);
    // svm gives distances, needed are probs
    results = 1.0f - results;

    std::vector<Point2f> roc;
    ROC::curve(results, vlabels, roc, 100);
    cerr << "AUC " << ROC::auc(roc) << endl;

    Mat roc_draw(480, 640, CV_8UC3, Scalar::all(255));
    ROC::draw(roc, roc_draw, Scalar(255,0,0));
    imshow("ROC",roc_draw);
    waitKey();

    return 0;
}