PhD Videos

Parametric Face Alignment: Generative and Discriminative Approaches

PhD Thesis - Videos

Pedro Martins
Institute of Systems and Robotics (ISR)
University of Coimbra, Portugal



# Chapter 2: Generative 2.5D Active Appearance Models


Efficient Simultaneous Forwards Additive (ESFA) Fitting Algorithm


Video 2.1: The 2.5D AAM fitting using the Efficient Simultaneous Forwards Additive (ESFA) algorithm. Each image show the input frame overlaid with the projected mesh and tree different views of the current 3D mesh s.


Efficient Robust Normalization Forwards Additive (ERNFA) Fitting Algorithm


Video 2.2: Robust 2.5D AAM fitting using the Efficient Robust Normalization Forwards Additive (ERNFA) algorithm. The weight mask is shown on the right. The scale parameters were estimated assuming that there always exists 10% of outliers.


Efficient Robust Simultaneous Forwards Additive (ERSFA) Fitting Algorithm


Video 2.3: Robust 2.5D AAM fitting using the Efficient Robust Simultaneous Forwards Additive (ERSFA) algorithm. The weight mask is shown on the bottom-right. The scale parameters were estimated from the fitting error MAD on the images of the training set. The black triangles of the face mesh are in occlusion.


Fitting on the 3D Dynamic Facial Expression Database (BU-4DFE) using ESFA Fitting Algorithm



Video 2.4: Some examples of BU-4DFE dataset used for the evaluation of the 3D recovered shape when using the 2.5D AAM. The BU-4DFE dataset includes high resolution 3D dense reconstructions of video sequences of several individuals showing the six prototypic facial expressions namely, anger, disgust, happiness, fear, sadness, and surprise. The 3D facial expressions were captured at 25 frames per second where each expression sequence contains about 100 frames (resolution of 1040x1392 per frame).


Fitting in the FGNET Talking Face Video Sequence using the ERSFA Algorithm



Video 2.5: Qualitative 3D shape recovery on the Talking Face video sequence using the Efficient Robust Simultaneous Forwards Additive (ERSFA) algorithm.


Pose Estimation in the FGNET Talking Face Video Sequence


Video 2.6: Pose estimation results on the Talking Face video sequence using the Efficient Robust Simultaneous Forwards Additive (ERSFA) algorithm. Roll, Pitch and Yaw angles (in degrees) and distance (in mm) to camera (norm of the translation components) are shown. The distance is shown up to a scale factor due to the unknown focal distance of the TF sequence.



# Chapter 3: Discriminative Bayesian Active Shape Models


DBASM-KDE Labeled Faces in the Wild


Video 3.1: DBASM-KDE qualitative fitting evaluation in the Labeled Faces in the Wild (LFW) dataset.


DBASM-KDE Fitting in the IMM Database


Video 3.2: Fitting at Image Frame (Left); Fitting at the BaseMesh (Center); Mean-shift paths + uncertainty covariances (Right).


DBASM-WPR FGNET Talking Face Evaluation


Video 3.3: DBASM - Weighted Peak Response (WPR) evaluation in the Talking Face Video sequence.


DBASM-GR FGNET Talking Face Evaluation


Video 3.4: DBASM - Gaussian Response (GR) evaluation in the Talking Face Video sequence.


DBASM-KDE FGNET Talking Face Evaluation


Video 3.5: DBASM - Kernel Density Estimator (KDE) evaluation in the Talking Face Video sequence.


BASM-KDE Labeled Faces in the Wild


Video 3.6: BASM-KDE qualitative fitting evaluation in the Labeled Faces in the Wild (LFW) dataset.


BASM FGNET Talking Face Evaluation


Video 3.7: BASM (WPR/GR/KDE) evaluation in the Talking Face Video sequence.



# Chapter 4: Identity and Facial Expression Recognition


Identity and Facial Expression Recognition


Video 4.1: Examples of identity and expression recognition (F4 Testing Fold). The left image show the AAM fitting, the expression trajectory on the manifold is represented as a black path at center image and the projected test point into the identity manifold is the black dot at right image.