False SURF output in video

I am trying to find the three cervical vertebrae (C2 , C3 , C4)in x-ray videos (gray images) I am using SURF to identify the place that has these 3 vertebrae (big box for all of them )
SURF gives me good result in this case

Now I am trying to identify each one of these vertebrae from this big box, but SURF failed in this case The SURF works very good with the area that has 3 vertebrae , but not good to find each of them separately

Update : The idea of SURF is finding the important keypoints in the image. therefore, I am using this idea to find the keypoints in 2 images then find the match points between these 2 images. in the beginning of my program , the user select 4 boxes. one big box for all vertebrae and the other 3 of each one.

I will use this selected area as a template to find them in the next images in the videos. first I will apply SURF to find the big box of the all vertebrae. and this work good with this code

now I am tying to find the three small boxes inside the big one but SURF gives me bad results (wrong boxes)

The photo show you the 4 boxes (ignore the left box and the three points ) these are the 4 template images that I am using (after crop it 4 times)

The question , How can I improve the SURF results to get the three vertebrae ?

any help will be so appreciate :D

here is the image that show the perfect results .... big box has 3 boxes ...

This is the code that I am using First parameters (Mat img_object) is the template image that I am trying to find in (Mat img_scene) , this second paramter is the big box that has the 3 vertebrae , third parameter is the size of the box that I wanna draw around the object when we find it

CvRect Identify_SURF_Frame (Mat img_object , Mat img_scene , CvRect in_box)
{
cvNamedWindow("Good Matches & Object detection", CV_WINDOW_AUTOSIZE);
CvRect output_box;
//-- Step 1: Detect the keypoints using SURF Detector
int minHessian = 1;  // I reduce this number so I can have a lot of number for  keypoints
SurfFeatureDetector detector( minHessian , 2 , 3 , true , true );
std::vector<KeyPoint> keypoints_object, keypoints_scene;
detector.detect( img_object, keypoints_object );
detector.detect( img_scene, keypoints_scene );

//-- Step 2: Calculate descriptors (feature vectors)
SurfDescriptorExtractor extractor;
Mat descriptors_object, descriptors_scene;
extractor.compute( img_object, keypoints_object, descriptors_object );
extractor.compute( img_scene, keypoints_scene, descriptors_scene );

//-- Step 3: Matching descriptor vectors using FLANN matcher
BruteForceMatcher < L2 < float > > matcher;
std::vector< DMatch > matches;
matcher.match( descriptors_object, descriptors_scene, matches );
double max_dist = 0; double min_dist = 100;

//-- Quick calculation of max and min distances between keypoints
for( int i = 0; i < descriptors_object.rows; i++ )
{
double dist = matches[i].distance;
if( dist < min_dist ) min_dist = dist;
if( dist > max_dist ) max_dist = dist;
}

//-- Draw only "good" matches (i.e. whose distance is less than 3*min_dist )
std::vector< DMatch > good_matches;
for( int i = 0; i < descriptors_object.rows; i++ )
{
if( matches[i].distance < 4 * min_dist )
{
good_matches.push_back( matches[i]);
}
}

Mat img_matches;
drawMatches( img_object, keypoints_object, img_scene, keypoints_scene, good_matches, img_matches, Scalar::all(-1), Scalar::all(-1),
vector<char>(), DrawMatchesFlags::NOT_DRAW_SINGLE_POINTS );

//-- Localize the object
std::vector<Point2f> obj;
std::vector<Point2f> scene;
if (good_matches.size() >= 4)
{
for( int i = 0; i < good_matches.size(); i++ )
{
//-- Get the keypoints from the good matches
obj.push_back( keypoints_object[ good_matches[i].queryIdx ].pt );
scene.push_back( keypoints_scene[ good_matches[i].trainIdx ].pt );
}

Mat H = findHomography( obj, scene, CV_RANSAC );
//-- Get the corners from the image_1 ( the object to be "detected" )
std::vector<Point2f> obj_corners(2);
obj_corners[0] = cvPoint(0,0);
obj_corners[1] = cvPoint( img_object.cols, 0 );
//obj_corners[2] = cvPoint( img_object.cols, img_object.rows );
//obj_corners[3] = cvPoint( 0, img_object.rows );

std::vector<Point2f> scene_corners(2);
perspectiveTransform( obj_corners, scene_corners, H);
int x1 , x2 , y1 , y2 ;
x1 = scene_corners[0].x + Point2f( img_object.cols, 0).x ;
y1 = scene_corners[0].y + Point2f( img_object.cols, 0).y ;
x2 = scene_corners[0].x + Point2f( img_object.cols, 0).x + in_box.width ;
y2 = scene_corners[0].y + Point2f( img_object.cols, 0).y + in_box.height ;

rectangle(img_matches , cvPoint(x1, y1) , cvPoint(x2, y2)  , Scalar( 255, 255, 255), 1 );
output_box.x = x1 - in_box.width ;
output_box.y = y1 ;
output_box.width = in_box.width ;
output_box.height = in_box.height ;
}
//-- Show detected matches
imshow( "Good Matches & Object detection", img_matches );
return output_box ;
}

• 1) Uploading images somewhere we can see them will really help. 2) In general, code is not a good way to start here on DSP.SE. Try explaining your algorithm in pseudo-code so that those not intimately familiar with the OpenCV library can understand (and hopefully contribute) too! (Like me!). – Peter K. Apr 29 '13 at 18:17
• I edit the post to have a photo. for explaining the code, I am using known algorithm called SURF. this will be too long to describe here – seereen Apr 29 '13 at 18:33
• I agree with what Peter said, and while I understand that you don't want/need to describe the entire SURF algorithm, you should give more non-code detail on your implementation method. For instance, rather than simply telling us that 'Mat img_object' is the template, show the template in the post. And try to include some actual results of your current implementation, not just the ideal results. – Sam Maloney Apr 29 '13 at 19:28

In brief, what you are trying to do is probably not feasible. Why? Because SURF was not designed to do that.

In a first approximation, SURF is designed as a faster (approximated and with less invariance) SIFT, i.e., a computer vision keypoint detector. Roughly speaking, CV keypoints are either corners or blobs as detected by some kind of difference of Gaussians (the scale-space search is used to auomatically select the best Gaussian widths).

While these keypoints are present in structure-from-motion or urban stereoscopic images, they are clearly absent in your images. If your primary goal is to detect the vertebrae, you should try instead (and for example):

• template matching approaches (since you have an accurate visual model of what you are looking for)
• or maybe train some cascade or HoG based detector that will be applied using sliding windows on your full images.
• why I can do it for the big box not for the small boxes. I think SURF cannot work with small images , right? – seereen Apr 30 '13 at 17:31
• the suggestion ideas need to do training, I am trying to find another ideas without training like SURF. I try SIFT and BRISK both of them do not give me any keypoints at all !!! – seereen Apr 30 '13 at 17:38
• I have to find something better than template matching :( – seereen Apr 30 '13 at 18:19
• BRISK, SIFT, FREAK, etc. are all corner-oriented. You have no corners at all, so template matching is still you best hope. Maybe linear feature extraction to detect the bones ? – sansuiso Apr 30 '13 at 20:05
• If I want to use Template matching ... how can I apply the rotation ??? ... the three Cervical rotate little bit in the videos??? – seereen May 6 '13 at 19:37