# How to generate key points by myself

Hi, i want to compute SURF descriptors on my own key points. The problem is, currently I only have 2d image points (x,y-coordinates), but opencv's method for calculating SURF descriptors needs keypoints, which need beside the coordinates additional information like, scale, size, orientation,... (information which I do not have).

Please help -> How can I generate correct Keypoints out of my 2d image points? (Correct means, I want to have key points on exactly the coordinates of my image points).

edit retag close merge delete

just curious, what are you trying to do with it ?

( 2016-06-16 06:39:19 -0500 )edit
1

I'm trying to train an SVM on these descriptors. The points are image points on a persons face (e.g eyes, mouth...) and to train an svm on this descriptors to later be able to do person identification i need exactly the same points on every training image (which i have) but then also the descriptors on this points. Hope this makes it understanable :-)

( 2016-06-17 03:50:28 -0500 )edit

yes, sure makes sense !

( 2016-06-17 04:27:35 -0500 )edit

Sort by » oldest newest most voted

the "minimal" args to construct a Keypoint are x,y,size. you'll just have to invent a size, that makes sense in your image (e.g avg. distance between points).

vector<Point> pts = .... ; // from landmarks or such
vector<KeyPoint> kp; // empty
float s = 5; // ??
for (size_t i=0; i<pts.size(); i++)
{
kp.push_back(pts[i].x, pts[i].y, s);
}

Mat descriptors;
Ptr<Feature2D> extractor = features2d::SURF::create(); // you need float features for the svm !
extractor->compute(image, kp, descriptors);

more

Thanks for your answer. I tried this before, but the recognition of the identity using the svm is super bad. A friend said, that he thinks it is because if I only use these three values (x,y,size) the response is default 0 and the angle is -1, that leads to keypoints which are not meaningful at all. I also tried to set the response to the same value for all keypoints and the angle to 0, but the results were really bad also :-(

( 2016-06-17 04:35:08 -0500 )edit

"but the recognition of the identity using the svm is super bad."

yea, that's kinda expected ;( true, your points are not meaningful for the descriptors. you could try AKAZE instead of SURF with DESCRIPTOR_KAZE_UPRIGHT as descriptortype (since it does not expect angles) , or a "dense" grid , say 5 pix apart, instead of your landmarks, also tweaking SVM params, but in the end ...

btw, how many landmarks do you have ? (and from where?)

( 2016-06-17 04:43:51 -0500 )edit

Thanks for that idea, I will give this a try. For the identification I use 11 landmarks (image points) i get them using the face-landmark-detector from dlib.

( 2016-06-17 05:30:20 -0500 )edit

Official site

GitHub

Wiki

Documentation