Face Analysis Under Uncontrolled Conditions

8Towards Adaptation to Head Pose Variations

Benjamin ALLAERT¹, Ioan Marius BILASCO² and Chaabane DJERABA²

¹IMT Nord Europe, Lille, France

²University of Lille, France

8.1. Introduction

Most facial expression analysis systems are designed to analyze faces under good acquisition conditions (face fixed and frontal to the camera). This ensures that the visual features are perfectly exploitable. However, in a natural context, where the acquisition conditions change, the faces are not directly exploitable on most current systems. This is mainly the case in the presence of pose variations (2D and 3D displacements of the face with respect to the camera), where the face tends to become partially obscured.

In view of recent approaches to characterize facial expressions, dynamic approaches are more suitable because they allow for the detection of subtle movements. However, the use of these approaches requires that dynamic textures be perfectly segmented locally and spatially in order to avoid inducing motion discontinuities. It is therefore important to ensure that the face is correctly aligned throughout the analysis sequence despite the presence of pose variations or large displacements to get the full benefit of these approaches.

The solution used to adapt dynamic approaches to pose variations and large displacements is to add a preprocessing step that consists of normalizing the face in order to bring the face to an ideal analysis configuration by correcting the geometric transformation ...