Automatic 3D Face Modeling Using 2D Active Appearance Models

Although a lot of promising research findings have been studied on 3D face modeling in the past years, it is still a challenge to generate realistic 3D human face models and facial animations. This paper presents a novel approach to model 3D faces automatically from still images or video sequences without manual interactions. Our proposed scheme comprises three steps. First, we offline construct 3D shape models using Active Appearance Models (AAMs), which saves large computation costs for online modeling. Second, based on the computed 3D shape models, we propose an efficient algorithm to estimate the parameters of 3D pose and non-rigid shape via local bundle adjustment. Third, we employ the recovered 3D face shape to deform the high resolution face mesh through scattered data interpolation, and extract the face texture maps for rendering the reconstructed face model from various viewpoints. Since the correspondence is built by AAMs fitting, the 3D face model can be constructed effectively even from a single image.