3D face modeling from image sequence

Abstract

Generation of realistic 3D human head models has a wide and promising range of applications in many fields like computer games, video conferencing, filmmaking, security, identification, healthcare, interactive media, etc. However, it still remains as a challenging task in computer graphics for many years. Presently, devices such as laser range scanners or 3D scanners offer accurate modelling for a wide array of complex objects, but they are expensive and time consuming. In my Final Year Project, I work on a system which can create a 3D face model from a sequence of 2D images capturing the actor’s face. The system first extracts and tracks face features from a sequence of images, and then it estimates the motion and structure parameters and reconstructs a specific 3D face by Extended Kalman Filter algorithm. Rao-Blackwellized Particle Filter algorithm for motion and structure estimation is also derived and developed in my work. In the next stage, due to the errors and lack of information in the non feature regions, the specific face model is combined with a generic face model using Monte Carlo Markov Chain framework. Then texture mapping is performed, which results in a grayscale 3D face model. Finally by color rendering, the system produces a colored 3D face model. My work mainly focuses on motion estimation and specific face model reconstruction as well as color rendering of the face model. The system and my work are fully described in this Final Year Project report, together with the results and discussions. Some recommendations for further implementations are also discussed at the end.