Head Pose Estimation AAM + POSIT

Example of Head Pose Estimation using 2D Active Appearance Models (AAM) combined with Pose from Orthography and Scaling with ITerations (POSIT). The 2D AAM (running the Project Out algorithm) fits each frame of the sequence and the 6 DOF are estimated using a 3D rigid model of the face with POSIT.

See details at http://www.isr.uc.pt/~pedromartins

Category:

Science & Technology

Tags:

• Active Appearance Models
• AAM
• POSIT

[dailydols]

It is great to see how tolerant POSIT can be.

Is it a RANSAC or a least-square approach in this POSIT implementation?

[pedroadmartins]

@dailydols

Hi. I'm using a LS implementation. With a lot of 2D/3D correspondences the POSIT performs very well (assuming that you have a good 2D feature tracker).

[couash]

Nice work. I'd like to play around and experiment with AAM, but I am not into computer vision nor do I have a lot of time to learn the theory weel enough to implement it myself. Does anyone know an open source implementation of AAM that could run on a mac? What did you use here? OpenCV? Any chance you'd share your code?

[dailydols]

In my continues studies I always end up in that video of yours again, it is just great :-)

[dailydols]

just found your paper "Monocular Head Pose Estimation".

good work :-)

[dailydols]

when you reach certain side angles you could retrain you model with the new textures to include changes in lighting to enhance it even further.

[dailydols]

I am using an adaptive method in a theses, maybe I record a video and post in on youtube.

[dailydols]

What you could do next is to take the simple 3D mesh and use a frontal image as a texture, rotate the textured mesh over a set of angles an use these images to train a new model which fits better on your face. when this is done you can track your face to lets say +- 45degrees.