[Home] / Pose-free Facial Landmark Fitting

Pose-free Facial Landmark Fitting via Optimized Part Mixtures
and Deformable Shape Model




Description

In this work, we present a novel framework to handle large pose variation in facial landmark localization and tracking. A group sparse learning method is proposed to automatically select the optimized anchor points. We set up weights for each landmark patch in the part mixture model indicating the likelihood of choosing these parts. By regularizing the weights group sparse, maximizing the margin over positive and negative training samples generates effective weights to simplify the mixtures of parts. By introducing 3D face shape model, we use procrustes analysis to achieve pose-free facial landmark initialization. Then a two-stage cascaded deformable shape model is applied to effectively and efficiently localize the facial landmarks. Validating the result on not only laboratory face databases but also face-in-the-wild databases, our framework reveals its advantages in handling extreme head pose variation, expression and complex background.

Publication

X. Yu, J. Huang, S. Zhang, W. Yan, D.N. Metaxas. "Pose-free Facial Landmark Fitting via Optimized Part Mixtures and Deformable Shape Model" 14th IEEE International Conference on Computer Vision(ICCV), Sydney, Australia, Dec. 2013. [PDF] [Supp] [Matlab Code] [BibTex]

Tracking Videos

1. Talking Face Video

2. Youtube Celebrity Videos

3. Other Videos