Modeling Age, Obesity, and Ethnicity in a
Computerized 3-D Facial Reconstruction

Paper presented at the 9th Biennial Meeting of the International Association
for Craniofacial Identification, FBI, Washington, DC, July 2000

Martin P. Evison
Department of Forensic Pathology
University of Sheffield
Sheffield, United Kingdom

Introduction.......Research at Sheffield.......Discussion.......References


Introduction

It is a common misconception that a facial reconstruction is an exact likeness of a person during life. A cursory examination of the skull reveals that the relationship between skull contour and facial appearance is not direct. On the forehead, margins of the eyes, cheekbones, bridge of the nose, above the lips, and the chin, facial shape is closely related to skull contour. However, the shape of the eyes and eyelids, the tip of the nose, and the lips cannot be predicted from the skull, and these are important features in facial recognition (Figure 1). Furthermore, hairstyle and hair color, skin color, facial hair, scars, tattoos, piercings, jewelry, clothing, cosmetics, glasses, and numerous other aspects of facial appearance cannot be determined from the skull. Estimates of age, height, build, and even sex are imprecise, and degrees of obesity and aging, for example, cannot be predicted. Ethnic affiliation can only be estimated with weak statistical confidence and even then only for that which has evolved as a consequence of continental scale biodiversity and history—such as African, African-Caribbean, or Caucasoid (European ancestry).

Figure 1. Photograph of a facial reconstruction in progress, illustrating that the shape of the eyes, tip of the nose, and lips are largely guesswork.

Many of the measurements used in facial reconstruction are tissue depths. These measurements were collected from cadavers. The individual in death bears limited resemblance to the individual during life. Postmortem changes and gravity affect tissue-depth measurements. The data sets are small. Measurements are imprecise and are taken from approximately 20 landmarks. There is no way of knowing when applying the mean tissue depths from a data set that it is a mean member of the population. A reasonable resemblance is the best that can be hoped for from facial reconstruction, and even this seems something of an achievement.

Traditional plastic reconstruction using clay is slow. In an emergency, a reconstruction can be completed in one day but will normally take up to a week. Reconstructions will vary even when repeated by the same practitioner and will differ widely between practitioners. There is a well-known tendency to incorporate one's own facial features into a reconstruction. Reconstruction relies on a substantial degree of dexterity and artistic skill, as well as on knowledge of anatomy and physical and dental anthropology. Amendments to the reconstruction to account for aging or obesity, for example, mean several hours' additional work in the studio, and if the original outcome is to be kept for posterity, a copy must be made. Further time constraints are introduced by the simple logistics of bringing skull and practitioner together in national and international investigations. Research at Sheffield A computer system for facial reconstruction is being developed at Sheffield University, Sheffield, United Kingdom (Evison 1996; Evison et al. 1998; Tyrrell et al. 1997). This system is based upon the ability to capture a 3-D digital image of a skull (Figure 2) using a laser scanner. The aim is to create a computer system that is accurate, rapid, repeatable, accessible, and flexible. Although accuracy in facial reconstruction has increased recently with the introduction of measurements taken from ultrasound scans of living individuals (Helmer 1984), these data sets are still based upon the approximately 20 landmarks traditionally used. The goal is to increase accuracy by means of tissue-depth data collection from living individuals using magnetic resonance imaging (MRI) equipment. MRI scans of the head are captured as a series of 2-D slices (Figure 3) stored in digital form as a stream of gray-scale values. The soft tissue of the face and the underlying bone contrast clearly. Figure 4 illustrates the output from an algorithm that generates separate 3-D surfaces of the soft tissue and bone from a series of MRI slices. Manual tissue-depth data collection is presently possible (Figure 5), and current research is directed at automating tissue-depth data collection. It may be feasible to collect about 100 tissue-depth measurements from 100 MRI slices, amounting to 10,000 tissue-depth measurements per head. The research has also attempted to improve the accessibility of facial reconstructions. The initial step was to develop simple protocols for placing plastic reconstructions in 3-D on the Internet. The preliminary models developed using Silicon Graphic's Open Inventor™ (Evison and Green 1998) were easily ported into Virtual Reality Modeling Language (VRML) for presentation on the Internet (Evison and Green 1999). In this way global access to facial reconstructions was quickly achieved. Once in VRML format, the reconstructions were enhanced using standard 3-D editing software. This allowed experimentation with point light sources, fixed viewpoints, skin-like textures, and facial features (Figure 6). Flexibility is also enhanced in VRML through the use of interpolators, which allow changes in shape and color to be built into the 3-D model. Interpolators were used to experiment with three of the major unknowns of facial reconstruction: obesity, aging, and ethnic affiliation. Prototype models were developed using scanned plastic reconstructions as standards for the ranges of obesity versus emaciated, young adult versus older adult, and Caucasoid versus Negroid (African ancestry). Using this approach, the interpolation model that morphs between the emaciated and obese values of the published tissue-depth data sets was developed. A range of possible outcomes representing the potential range of obesity is represented in a 3-D model (Figure 7). A white skin-like texture was applied, and three fixed viewpoints that led the viewer around the model were also included (Green and Evison 1999). Although tissue-depth data is available that can be used to model obesity, no measurements exist that can be used to model the aging process in adults. Published data on aging tends to be qualitative rather than quantitative and is largely descriptive. This information was used to model aging of standard reconstructions taken from typical male and female skulls.

Figure 2. Wire-frame representation of a digitized 3-D skull image. Figure 3. A magnetic resonance imaging (MRI) slice of the head in which the bone and soft tissues of the face are clearly distinguished. Figure 4. Cross-section image of 3-D soft tissue and bone surfaces of the head of a three-year-old child generated from a series of 2-D MRI slices. Figure 5. Manual tissue-depth data collection from a 3-D image of soft tissue and bone surfaces. Figure 6. Using graphical editing software to incorporate a point light source and a moustache.

Figure 7. Interpolating between obese and emaciated facial reconstructions. Available: http://forensic.shef.ac.uk/fatten.wrl

Representations of the reconstructed face aged to approximately 30, 50, and 80 years were produced. Scans of these plastic models were used to develop interpolation models of aging in VRML. Simple linear interpolation models morphing between the 30- and 80-year-old age standards were developed first. It is important to recall that the aging process is nonlinear. Aging will progress at different rates for different people and will affect different parts of the face in different ways. In order to explore the potential for the complex simulation of aging, a nonlinear interpolation model of aging was developed, progressing from 30 to 80 years using a 50-year fixed standard (Figure 8).

Figure 8. Linear interpolation between young and old facial reconstructions of a male. Available: http://forensic.shef.ac.uk/ageing.wrl

It is not possible to divide humanity into biological races on scientific grounds. Nearly all human genes are shared by all human populations, as well as with chimpanzees and gorillas. The range of genetic variation within any ethnic group substantially exceeds the net difference between any two groups. There are no genetic variants possessed entirely and exclusively by any single ethnic group. Races in the biological sense do not exist.

Nevertheless, there are a small number of biometric characters that are significantly more frequent in one ethnic group than in another. Models based upon biogeographical history can explain these differences. They are largely trivial, but in a forensic context they can be used to estimate whether the individual would have been likely to be identified with one or another ethnic group. It seems obvious that this can be important in establishing the identity of an individual. Estimation of ethnic affiliation is highly inaccurate, however, because of the substantial overlap in variation between ethnic groups, even those historically separated by considerable geographical distance.

In order to examine how ethnicity-related variation in facial appearance might be simulated in computerized 3-D models, casts of an archaeological skull that contained features more frequent in the Caucasoid population along with features more frequent in the Negroid population were made. This person could have been Caucasoid, Negroid, of mixed ancestry, or one of a number of other potential ethnic affiliations. Standard reconstructions based upon the published tissue depths collected from samples of Negroid and Caucasoid populations were produced. Scans of these reconstructions were used to construct a VRML model that interpolates between the Caucasoid and Negroid reconstructions made from the same skull. A color interpolator node was also included so that the skin color hue changed in correspondence with the tissue-depth data being applied (Figure 9). It is important to remember that light skin color need not accompany Caucasoid facial features in an individual of mixed African and European ancestry and that these features may also be inherited in discrete non-additive ways. Nevertheless, the model demonstrates that computer simulations can be used to offer a range of variation from which certain outcomes of reconstructions can be selected in such cases.

Figure 9. Interpolating between the Caucasoid and Negroid tissue depths. Available: http://forensic.shef.ac.uk/ethnic.wrl

Discussion

The research and models are preliminary in nature. Abandoning the traditional landmarks in favor of approximately 10,000 measurements collected from MRI records will represent a significant increase in the precision and accuracy of facial reconstructions. However, difficulties in automating the process by which the tissue-depth landmarks are located on the skull using the traditional landmarks will be encountered.

There is need for substantial research on the simulation of skin color and texture and on the use of 3-D graphical editing software to enhance the finished reconstruction. Although the development of aging and other interpolation models represents a significant advance, the models presently lack the sophistication and accuracy of life.

Little work has been undertaken examining the value of studies in psychology, art history, or portraiture in facial reconstruction. As a field that straddles the humanities, arts, and sciences, facial reconstruction offers a rare opportunity for interdisciplinary research.

The forensic applications of computerized 3-D reconstruction are noteworthy. Computerized reconstructions should substantially reduce the cost of facial reconstructions to investigating authorities, although hard copies are likely to remain costly for the time being. An international service for computerized facial reconstruction via the Internet is now feasible, and it will be possible to use portable computerized facial reconstruction systems in human rights abuse investigations in the field.

References

Evison, M. P. Computerized 3D facial reconstruction, Assemblage [Online]. (1996). Available: http://www.shef.ac.uk/~assem/1

Evison, M. P. and Green, M. A. Computerized facial reconstruction using Open Inventor™. In: Let's Face it!: Proceedings of the 7th Meeting of the International Association for Craniofacial Identification. International Association of Craniofacial Identification, University of Melbourne, Melbourne, Australia, 1998, pp. 43–46.

Evison, M. P. and Green, M. A. Presenting three-dimensional forensic facial simulations on the Internet using VRML, Journal of Forensic Sciences (1999) 44:1216–1220.

Evison, M. P., Finegan, O. M., and Blythe, T. C. Computerized 3-D facial reconstruction: Research update, Assemblage [Online]. (1998). Available: http://www.shef.ac.uk/~assem/4

Green, M. A. and Evison, M. P. Interpolating between computerized three-dimensional forensic facial simulations, Journal of Forensic Sciences (1999) 44:1221–1225.

Helmer, R. Sch@delidentifizierung durch Electronische Bildmischung. Kriminalistik Verlag, Heidelberg, 1984.

Tyrrell, A. J., Evison, M. P., Chamberlain, A. T., and Green, M. A. Forensic three-dimensional facial reconstruction: Historical review and contemporary developments, Journal of Forensic Sciences (1997) 42:653–661.

Top of the page