Ask Your Question

FREAK experience requiered

asked 2015-03-16 03:38:00 -0500

ArcadeBit gravatar image

Hello there,

I am working on a facetracking for quite a while now, to control a virtual 3D scene. I run a simple faceDetection on an 320x240 image and track the position with the detected face. Sadly, the face detection isn't rotation invariant, so I createt a new classifier, but the false positive rate was just terrible. So I started to rotate the image (20 degree to the left and right), detected my Face on 3 images parralel in separate Threads, rotate my Keypoint (if requiered) back and done. This kinda works, but the Thread handling isn't nice.

My solutuion is kinda fast. I get my Keypoint in under 10ms, but I don't like the solution. Now I wanted to give FREAK, SURF, SIFT and ORB a chance. How fast are these OpenCV modules? Can anyone tell me a realistic time to detect a face Position in a 320x240 image?

edit retag flag offensive close merge delete


You can try it...

thdrksdfthmn gravatar imagethdrksdfthmn ( 2015-03-16 03:48:35 -0500 )edit

1 answer

Sort by ยป oldest newest most voted

answered 2015-03-16 13:11:32 -0500

updated 2015-03-16 13:12:15 -0500

FAST turned out to be the fastest on my Haswell machine on a 640x480 rgb image from the Kinect, but I think this could be dependent on the scene ( mine is a top down scene with just people in it ). Valid matches is again a whole other story. I am afraid you will have to test them all.

Are you planning on finding the face first and then track the features that you find in the area? otherwise I am not too sure if you will be able to match faces from a set of features found in a set of images of faces.

edit flag offensive delete link more


By the way I had bad matching results when using FREAK as a descriptor

Christoph Pacher gravatar imageChristoph Pacher ( 2015-03-17 11:55:00 -0500 )edit

I want to detect the face with the face-detector first and then switch to tracking with FAST, FREAK, SURF, SIFT or BRISK, depending on the performance (speed and accuracy). Can you tell me, how fast your solution was?

ArcadeBit gravatar imageArcadeBit ( 2015-03-17 12:41:11 -0500 )edit

just to give you a point of reference, i find with FAST/BRISK on an i5 Haswell processor about 150 kps per image and match them to the previous image in about 2ms. this is highly dependent on the image and the number of kps you find/match. Most of the other Matcher/Descriptor combinations took longer in my tests, up to 25ms.

Christoph Pacher gravatar imageChristoph Pacher ( 2015-03-21 14:26:45 -0500 )edit

Question Tools

1 follower


Asked: 2015-03-16 03:38:00 -0500

Seen: 359 times

Last updated: Mar 16 '15