Research Article: How Well Do Computer-Generated Faces Tap Face Expertise?

Date Published: November 4, 2015

Publisher: Public Library of Science

Author(s): Kate Crookes, Louise Ewing, Ju-dith Gildenhuys, Nadine Kloth, William G. Hayward, Matt Oxner, Stephen Pond, Gillian Rhodes, Alexandra Key.


The use of computer-generated (CG) stimuli in face processing research is proliferating due to the ease with which faces can be generated, standardised and manipulated. However there has been surprisingly little research into whether CG faces are processed in the same way as photographs of real faces. The present study assessed how well CG faces tap face identity expertise by investigating whether two indicators of face expertise are reduced for CG faces when compared to face photographs. These indicators were accuracy for identification of own-race faces and the other-race effect (ORE)–the well-established finding that own-race faces are recognised more accurately than other-race faces. In Experiment 1 Caucasian and Asian participants completed a recognition memory task for own- and other-race real and CG faces. Overall accuracy for own-race faces was dramatically reduced for CG compared to real faces and the ORE was significantly and substantially attenuated for CG faces. Experiment 2 investigated perceptual discrimination for own- and other-race real and CG faces with Caucasian and Asian participants. Here again, accuracy for own-race faces was significantly reduced for CG compared to real faces. However the ORE was not affected by format. Together these results signal that CG faces of the type tested here do not fully tap face expertise. Technological advancement may, in the future, produce CG faces that are equivalent to real photographs. Until then caution is advised when interpreting results obtained using CG faces.

Partial Text

Advances in technology have seen an increase in the use of computer-generated (CG) stimuli in face processing research in recent years. Artificial faces with a very human-like appearance can now be generated by a number of software programs with ease (either ‘from scratch’ or by inputting real photographs to be converted into 3-D head models). Different facial characteristics can be specified or varied when generating these faces including sex, age, ethnicity and attractiveness. Once generated, the faces can then be easily manipulated for facial expression and viewpoint. CG faces are also highly standardised in terms of lighting conditions, extra-facial information, size and image quality. All these factors make CG faces very appealing to face processing researchers, particularly given the limitations that existing databases of face photographs often impose on experimental design and the cost and time required to generate new photographic databases. However little is known about the validity of the CG faces being used in research, and it remains unclear whether, as stimuli, they are equivalent to photographs of real faces.

In Experiment 1 we tested old/new recognition memory for Caucasian and Asian faces presented in three formats: Real face photographs, CG-Real (CGR) faces and CG-Artificial (CGA) faces. The two CG formats were chosen because they represent the two types of CG faces that have been used in previous studies. CGA faces were randomly generated by the software. This type of CG face is the most common in the literature (e.g., [2, 19, 26–28]). CGR faces were generated by importing the photographs from the Real condition into the software to produce CG versions of the Real faces, (e.g., [4, 25]). Including both CG formats provides a thorough test of the usefulness of CG face stimuli. CGR faces may also provide a fairer comparison to the Real faces than the arbitrarily generated CGA faces. Assuming 100% fidelity in the conversion process the CGR faces should be matched to the Real faces for within-set heterogeneity. As can be seen in Fig 1 the CGR faces retain some of the imperfections of the Real faces, but still lack some fine-grained texture information and may give a weaker impression of animacy. Texture information was not applied to the CGA faces as this has not been routinely done in previous studies. The CGA faces therefore have a uniformly smooth appearance.

In Experiment 2 we investigated whether the reduced own-race accuracy and reduced ORE for CG faces observed in Experiment 1 are restricted to recognition memory or also extend to perceptual discrimination. We used a simultaneous matching task in which participants had to match a target presented at the top of the screen to the same face identity in an array of 10 faces presented below the target [31]. On half the trials the target was not present in the array. A perceptual matching task including target absent trials was used to increase the difficulty of the task and because this task yields a clear ORE [22].

The results of this study provide important, new evidence that CG faces do not allow participants to demonstrate the full extent of their face expertise. This conclusion is based on the finding that two key markers of face expertise were diminished for CG compared to Real faces. In Experiment 1 recognition memory accuracy was significantly poorer for CG than Real own-race faces. The ORE was also significantly reduced or eliminated for CG compared to Real faces in three out of four conditions. In Experiment 2 perceptual matching accuracy on target absent trials was significantly worse for CG compared to Real own-race faces. Here, however, no differences between formats were found for the ORE. In combination these results suggest caution should be applied when using CG faces to examine expert processing of face identity.