1Anthropology, University of New Mexico, 2Maxwell Museum of Anthropology, University of New Mexico, 3Computer Science, University of New Mexico, 4, Google, Inc.
Saturday 55, Clinch Concourse
In New Mexico it is common to hear that there are two groups of Hispanics in the state, those primarily of Spanish descent and more recent immigrants of Mexican descent. Many New Mexicans say they can distinguish these groups by their appearance, including skin color, face shape, and hair form.
In an open-ended question, we asked 418 New Mexicans of Spanish-speaking descent to describe their sub-ethnic identity. Fifty-three distinct descriptors were used. We took three-dimensional facial photographs of these subjects, removing any data associated with coloration or shape below the mid-neck. Using unsupervised computer learning, we used a piecewise polynomial model of surface patches to represent facial structures with each patch representing a region of the face. We applied a genetic algorithm to select patches, and regularized least-squares to fit patches to face-scans. We use k-means clustering to group subjects in the polynomial coefficients, which represent the clusters of faces found to be similar.
We tested fit at 2-20 spline-plate patches and k=2-50 clusters, and found that 10 patches and 40 clusters best captured the variation among the faces, with only 0.05 mm loss per cluster.
Comparing the results of this unsupervised test to the subject’s own assessment of their ethnicity shows various levels of inconsistency among the groups. This may result from a lack of fit between self-identity and appearance, unequal distinctiveness of ethnic labels, or errors in the model. However, even with these caveats, it seems unlikely that two distinct groups can be identified by face shape alone.
This research was supported by the NSF, grant BCS 0962825.