3D Facial Recognition

There is something compelling about the idea that a face might be recognised not simply as a picture, but as a three-dimensional form. A photograph is always at the mercy of illumination and viewpoint: a familiar face can look surprisingly different when the light moves or the head turns. If, instead, we could recover the actual surface of the face—perhaps by projecting a pattern of light onto it and measuring how that pattern deforms—we might begin to describe the face in terms of its true geometry. The ridge of the nose, the depth of the eye sockets, the curvature of the cheeks: these are physical structures, not merely variations in image brightness.

The challenge, of course, is deciding what in that geometry constitutes identity. A 3D scan may contain thousands of surface points, but recognition cannot simply mean comparing one point cloud with another. The faces must first be aligned, perhaps using prominent landmarks or surface-registration methods such as Iterative Closest Point, and then reduced to some meaningful description. Distances, curvatures and relative proportions may prove useful, while statistical methods such as Principal Component Analysis may allow us to describe much of a face using a relatively small number of variables. It is an intriguing possibility that a person’s identity may be encoded not in any single feature, but in the overall pattern of shape.

If this can be made to work reliably, the implications are considerable. A three-dimensional representation should be less sensitive to changes in lighting and may allow us to compensate for variations in head pose by rotating the recovered surface into a common orientation. There remain serious practical problems—facial expression, missing data, acquisition speed and the cost of scanning equipment among them—but these feel like engineering obstacles rather than reasons to abandon the idea. One can imagine a time when the geometry of a face becomes a biometric signature in much the same way as a fingerprint. That possibility is scientifically exciting, although it also prompts an important question: if a face can be measured, stored and recognised automatically, we must also decide where such a capability ought to be used.