Next: Synthesizing 2D Images with Up: Inverse 3D Projection Previous: Occlusion and Back Surfaces

Mirroring Along the Mid-Line

In most scenarios, faces rotate along their vertical axis. For example, as people walk around in a room, they are unlikely to look downwards at the floor or upwards at the ceiling. They walk erect and do not tilt their heads excessively. The faces do, however, rotate along the vertical axis as people look left or right and walk in different directions. Thus, the most likely source of self-occlusion will be due to the rotation around the vertical axis. This will cause parts of the cheek or a side of the face to be occluded by the nose and the other side of the face. Luckily, the human face is symmetric across its vertical axis and the occluded part of the face closely resembles the visible part under vertical-axis rotations. Consequently, if the face is rotated along its vertical axis and the left or right half are not visible, we can take advantage of this symmetry to 'guess' what the hidden side of the face looks like. In situations where one side is hidden from the camera, we implement mirroring by copying the pixel intensities from the closer side of the face (the one most visible to the observer) onto the hidden side of the face as shown in Figure . Thus, mirroring is only implemented if the individual turns to the left or right and one side of the face is occluding the other. The arrows in the figure show which pixel intensities are mapped to which destinations to generate the mirror-symmetry in the human face.

Mirroring is only used if the nose is not well centered between the eyes (indicating a strong vertical axis rotation). Figure  displays the exact range in which the mirroring process is triggered. If the nose falls within the central strip between the two eyes, we do not rely on mirroring and simply assume no significant self-occlusion has occurred.

Next: Synthesizing 2D Images with Up: Inverse 3D Projection Previous: Occlusion and Back Surfaces
Tony Jebara
2000-06-23