Editing faces in videos

Editing faces in movies is of interest in the special effects industry. We aim at producing effects such as the addition of accessories interacting correctly with the face or replacing the face of a stuntman with the face of the main actor. The system introduced in this thesis is based on a 3D generative face model. Using a 3D model makes it possible to edit the face in the semantic space of pose, expression, and identity instead of pixel space, and due to its 3D nature allows a modelling of the light interaction. In our system we first reconstruct the 3D face, which is deforming because of expressions and speech, the lighting, and the camera in all frames of a monocular input video. The face is then edited by substituting expressions or identities with those of another video sequence or by adding virtual objects into the scene. The manipulated 3D scene is rendered back into the original video, correctly simulating the interaction of the light with the deformed face and virtual objects. We describe all steps necessary to build and apply the system. This includes registration of training faces to learn a generative face model, semi-automatic annotation of the input video, fitting of the face model to the input video, editing of the fit, and rendering of the resulting scene. While describing the application we introduce a host of new methods, each of which is of interest on its own. We start with a new method to register 3D face scans to use as training data for the face model. For video preprocessing a new interest point tracking and 2D Active Appearance Model fitting technique is proposed. For robust fitting we introduce background modelling, model-based stereo techniques, and a more accurate light model.