Given a collection of images, Pic1 analyzes them with k (currently over 8,500) extremely simple microfeature analyzers. Each analyzer produces a number in a certain range for each image. Thus, the k analyzers map the images to a -dimensional vector space of points. This gives Pic1 a rough idea of neighborhoods, nearness and farness, and clustering--in other words, it turns the collection of images into a metric space, whose basis set is some set of orthogonal images with respect to the k microfeatures.
Initially, Pic1 displays a group of n (currently 12) images randomly chosen from the database and says `pick one.' The user then picks the one that's most like, or least unlike, the one in mind. Pic1 then takes the vector representing that picture and generates n random points near to it in the metric space. It then displays their associated pictures and says `pick one' again. This repeats until the user has, with Pic1's aid, narrowed the search enough to find the image searched for.
The images may be photographs or may be drawings. Pic1 doesn't try to `understand' the images in any nontrivial fashion. It simply decomposes them and displays them, letting the user's selection guide it to find the searched for image. At no time, except perhaps for the initial analysis phase, has Pic1 any idea that the things associated with the vectors are images. So this scheme could potentially be used for other sorts of data that can be easily decomposed into a large number of dimensions of variation. There is a great deal of information given in just one selection of one multidimensional thing out of 12 multidimensional things.
At heart, Pic1 helps the user solve the search problem by imposing a notion of distance on the databases. This is often missing from most databases (Salton, 1991). Lacking distance metrics, most database queries have to be exact. Essentially, a user has to already know the database and how it is organized to use most other methods. Existing databases with distance metrics (e.g. Motro, 1988) generally exploit similarities in specific fields to permit approximate matches, but still require explicit queries.
Pic1 exploits what are perhaps the two best sides of each of the participants. It exploits the user's ability to quickly and easily extract information from large-detail pictures and it exploits the computer's ability to quickly do long error-free chains of calculations on large data sets. The system functions without forcing computers to act like people or forcing people to act like computers. We see this as a good fit between what humans are good at (fast image recognition and comparison) and what computers are good at (fast numerical computation).