I want to estimate head pose as part of my work. I have videos and i dont have frontal face pose, so i cannot have this constraint. It would be great to know your ideas how can i progress with this. Thanks
Belhassen, I read your paper, what i am not getting is when calculating YAW angle you use projections v1 and v2.. Are they image coordinates(x,y) of point on image plane? If yes, how do you calculate the difference? and i have a doubt if equation for calculating Roll angle is correct. I think it should be arccos(H/A).. If i am getting things right? and value of 'D' must be average approximation of distance of mouth corner to midpoint? Please comment on these issues.. thanks and Awaits!!
You can use Active Appearance Models. It gives you the tracked facial points together with the head pose. There are many implemented codes in C++ or in matlab...You would have to train your model on some of the faces you have in your database.
Another option is this : Face Detection, Pose Estimation, and Landmark Localization in the Wild..This is a very strong detector. You can find the implementation and the paper on: http://www.ics.uci.edu/~xzhu/face/
It depends on what you are doing: pose detection or pose tracking. Also, what camera you are using: monocular or stereo. Can you detect some facial feature points?
Thanks anam for your paper but my objective is not to find the head but the pose.. This paper is trying to detect head based on adaboost. It might not help in my particular scenario. I will try to explain my question again in the next comment..
@Hanan Salam and @Ngoc-Trung Tran please look at elaboration:
Problem Elaboration : I have a FACE VIDEO(s) of drivers in a car already taken from camera, I do not have intrinsic or extrinsic parameters of camera, maximum i can do is approximate the average distance between mouth and nose and eyes.. I have 68 landmarks on face, so i can use them to estimate pose, if its possible? I want to correct the pose to use for my application. Thanks.
So, I think one possible solution is to use 3D models, i.e. Candide-3 aligning to 2D landmarks using POSIT to compute 3D pose. If you do not know the intrinsic parameters, you can assume your own one. For example, focal length = 1000, and center of image is the projection center to build intrinsic camera matrice.