Fig. 2.

An example of using our uncalibrated multi-view fitting algorithm to fit a single 2D+3D AAM to three simultaneous images of a face. Each image is overlaid with the corresponding 2D shape for that image in dark dots. The head pose (extracted from the camera matrix ${\mathbf{P}}_{so}^N$) is displayed in the top left of each image as roll, pitch and yaw. The single 3D shape p̄ for the current ‘3-frame’ is displayed in the top right of the center image. This 3D shape is also overlaid in each image, using the corresponding ${\mathbf{P}}_{so}^N$, as a white mesh. See the movie iterations.mpg for a video of the whole fitting sequence