Exact solutions to Bayesian and maximum likelihood problems in facial identification when population and error distributions are known.

The reliability of traditional photogrammetric identification techniques using a small number of facial landmarks has recently come in for criticism. However, the transformation of parameters into a new face space in which the error distributions are orthogonal, yields a maximum likelihood solution to the problem of identifying a photographed face from a small, known, population which, in a simulated example, raises the success rate from 20% to 93%. A full transformation yielding simultaneously independent population and error distributions can be derived from raw population and error data using a straightforward computer procedure. Such a transformation facilitates computations for the situation where a single suspect is held in custody and the likelihood ratio of his being identical with a photograph is desired. It seems premature to condemn photogrammetry until the more efficient data-analysis approach outlined in this paper has been applied and tested.