Chernoff (Chernoff 1973, Chernoff & Rizvi 1975) presented an innovative method for creating mappings between computer-generated cartoon faces and points in a vector space. Chernoff wanted to use images (which people are good at processing) to understand and process vectors (which people are not as good at processing). Pic1 uses a similar mapping the other way, to let computers work with vectors (which computers are good at processing) instead of images.
Pic1's search process proceeds by presentation of examples from the database rather than by specifying abstract qualities about desired records; this is similar to RABBIT (Williams, 1984), which uses retrieval by reformulation to teach the user about the classification structure of the database. RABBIT, however, does not work with images because it is dependent upon the user clearly specifying which text fields are unacceptable for its reformulation; image perception is often holistic and thus not easily decomposed in this fashion.
The police sketch artist faces such a holistic problem. A witness has a vague recollection of a suspect's appearance, but has no good way to formalize the individual components in a query format. Caldwell and Johnston (Caldwell & Johnston, 1991) address this problem by asking the user to rate how closely each of several candidate images resembles the suspect; the user acts as the fitness function for a genetic algorithm (Goldberg, 1989). Such a system is still tedious to use, however, since the user must typically rank several hundred faces before convergence.
Some work has been done on addressing this problem with regard to faces; the CAFIIR system (Wu & Narasimhalu, 1994), for example, organizes images of faces through feature analysis and permits flexible retrieval in a variety of methods, including query-by-example and evolutionary browsing. But to date, systems have not typically attempted to generalize beyond restricted domains like frontal views of faces.