Eye detection in Oscar selfie - c++

Question

I have to perform eye detection on each of the faces in the famous Oscar selfie image.I tried using Haar Casacades on the faces since most of them are near-frontal, but the eye detection is totally random and no eyes are being recognized at all.
I have have tried the same haar cascade xml file for eye detection on images with single faces and it worked fine.
What steps could I take to correctly detect the eyes?
The image I used for eye detection can be downloaded from here:
https://drive.google.com/file/d/0B3jt6sHgpxO-d1plUjg5eU5udW8/view?usp=sharing
Below is the code I have written for face and eye detection. Basic idea is I first detect the face using viola jones algorithm and within each face, I try to detect the eyes.
#include <opencv2/highgui/highgui.hpp>
#include <cv.h>
#include <opencv2/objdetect/objdetect.hpp>
#include <vector>
using namespace cv;
using namespace std;
int x,y,w,h;
int main(int argc, const char** argv)
{
Mat image = imread("oscarSelfie.jpg",CV_LOAD_IMAGE_UNCHANGED);
Mat gray_img;
cvtColor(image, gray_img, CV_BGR2GRAY);
string faceCascade_file = "haarcascade_frontalface_alt2.xml";
string eyeCascade_file = "haarcascade_eye.xml";
CascadeClassifier faceCascade;
CascadeClassifier eyeCascade;
//Cascade classifier is a class which has a method to load the classifier from file
if( !faceCascade.load( faceCascade_file ) )
{ cout<<"--(!)Error loading\n"; return -1; };
//If it returns zero, it means an error has occured in loading the classifier
if( !eyeCascade.load( eyeCascade_file ) )
{ cout<<"--(!)Error loading\n"; return -1; };
equalizeHist(gray_img, gray_img);
//Increases contrast and make image more distingushable
/***** Detecting Faces in Image *******/
vector<Rect> faces;
vector<Rect> eyes;
//Rect is a class handling the rectangle datatypes
faceCascade.detectMultiScale(gray_img, faces, 1.1, 1, 0|CV_HAAR_SCALE_IMAGE, Size(30, 30) );
//faces.size()-it will return number of faces detected
for( int i = 0; i < faces.size(); i++ )
{
x = faces[i].x;
y = faces[i].y;
w = faces[i].width;
h = faces[i].height;
//Point center( faces[i].x + faces[i].width*0.5, faces[i].y + faces[i].height*0.5 );
//ellipse( image, center, Size( faces[i].width*0.5, faces[i].height*0.5), 0, 0, 360, Scalar( 255, 0, 255 ), 4, 8, 0 );
rectangle(image, cvPoint(x,y), cvPoint(x+w,y+h), CV_RGB(0,255,0), 2, 8 );
/******** Detecting eyes ***********/
eyeCascade.detectMultiScale(gray_img, eyes, 1.1, 50, 0|CV_HAAR_SCALE_IMAGE, Size(30, 30) );
for(int j=0; j < eyes.size(); j++)
{
Point center( faces[i].x + eyes[j].x + eyes[j].width*0.5, faces[i].y + eyes[j].y + eyes[j].height*0.5 );
int radius = cvRound( (eyes[j].width + eyes[j].height)*0.25 );
circle( image, center, radius, Scalar( 255, 0, 0 ), 4, 8, 0 );
}
}
namedWindow("oscarSelfie :)", CV_WINDOW_AUTOSIZE);
imshow("oscarSelfie :)", image);
waitKey(0);
destroyWindow("pic");
return 0;
}
`

Answer 1

i get the following result with facedetect.cpp (uses haarcascade_eye_tree_eyeglasses.xml)
don't expect to find all faces and eyes
i also tried dlib's face_landmark_detection_ex.cpp to compare results
dlib has an extra feature that gives you aligned faces like seen below

Answer 2

You may want to use CLM-framework for face landmark detection. As far as I have experience CLM-framework performance satisfactory.
Some examples of the system in action: http://youtu.be/V7rV0uy7heQ