Validation

Any judgment as to the validity of a facial model must be made in the context of the desired application. (Clearly the concept of ``accuracy'' is difficult to apply to an abstract cartoon representation of a face.) Validation processes are divisible into two broad classes: user-based and quantitative. User-based validation refers to asking the question ``does it look right?'' and in particular asking that question in the context of the particular application domain. For many applications of facial animation, this is the only feasible method of validation. Unfortunately, user-based validation does not always provide details about what is and isn't right about the model, and thus does not indicate a clear direction for improving the model.

The quantitative approach to validation can be employed in cases where the phenomenon being modeled can be measured. For facial animation of real people, there are several sources of data through which this could be done.

• Anthropometry data provides a database of measurements of the human body over the spectrum of race, gender, and age. In particular, cephalometrics is the study of measurements of the face. Cephalometric data can be used to determine in an average sense if the proportions in the model are ``natural'' for a human.

• Extensiometers are tools for measuring the relationship between load and displacement in tissue. For a physics-based facial model, a corresponding experiment may be done in the computer in order to provide a direct measure of the accuracy of the model.

• Dimensional Studies of Movement, as described by Ekman, provide a basis for comparing the movement of muscle groups in the model to normal values observed in a range of human subjects.

• Physiological Studies provide data about the characteristics of human behaviors in a form that may be cross-checked with the behavior of the model. For instance, the force and speed profiles of muscle movements have been studied in detail, as have the acceleration characteristics of the eye during saccades and visual pursuit. Speech behaviors have been particularly well documented.

• Optical tests provide a basis for validating the rendering performance of a model. Photographs of human faces taken under various lighting conditions could be simulated in the computer model and used to evaluate the shading model used to render the skin.