Generation of a 3-D face model from one camera

The generation of a fully textured 3-D model of a person's face presents difficult technical challenges, but has many applications in several fields, such as video games, immersive telepresence, and medicine. Current commercial systems rely on booth-like set-ups, equipped with laser-based scanners, or project a pattern on the subject's face. The major drawbacks of such systems are the cost of the hardware they require, and the lack of operational flexibility. We present here a fully automatic system to generate a 3-D model from a sequence of images taken by a single camera. Unlike other methods, we do not use a generic 3-D face subject to deformation, but instead proceed in a fully bottom-up fashion. The approach is a two-stage process. First, we estimate for each view the pose of the object with respect to the camera. This is accomplished by robust feature matching and global bundle adjustment. Then, we consider sets of adjacent views, which we treat as stereo pairs, and generate partial depth maps, which are then integrated into a single 3-D model. The texture is obtained by merging the images themselves. We describe the algorithm in detail, and show results on a number of real datasets.