Resumen de Spectrum analysis methods for 3d facial expression recognition and head pose estimation
Facial analysis has attracted considerable research efforts over the last decades, with a growing interest in improving the interaction and cooperation between people and computers. This makes it necessary that automatic systems are able to react to things such as the head movements of a user or his/her emotions. Further, this should be done accurately and in unconstrained environments, which highlights the need for algorithms that can take full advantage of 3D data. These systems could be useful in multiple domains such as human-computer interaction, tutoring, interviewing, health-care, marketing etc. In this thesis, we focus on two aspects of facial analysis: expression recognition and head pose estimation. In both cases, we specifically target the use of 3D data and present contributions that aim to identify meaningful representations of the facial geometry based on spectral decomposition methods:
1. We propose a spectral representation framework for facial expression recognition using exclusively 3D geometry, which allows a complete description of the underlying surface that can be further tuned to the desired level of detail. It is based on the decomposition of local surface patches in their spatial frequency components, much like a Fourier transform, which are related to intrinsic characteristics of the surface. We propose the use of Graph Laplacian Features (GLFs), which result from the projection of local surface patches into a common basis obtained from the Graph Laplacian eigenspace. The proposed approach is tested in terms of expression and Action Unit recognition and results confirm that the proposed GLFs produce state-of-the-art recognition rates.
2.We propose an approach for head pose estimation that allows modeling the underlying manifold that results from general rotations in 3D. We start by building a fully-automatic system based on the combination of landmark detection and dictionary-based features, which obtained the best results in the FG2017 Head Pose Estimation Challenge. Then, we use tensor representation and higher order singular value decomposition to separate the subspaces that correspond to each rotation factor and show that each of them has a clear structure that can be modeled with trigonometric functions. Such representation provides a deep understanding of data behavior, and can be used to further improve the estimation of the head pose angles.
