Searching images by content, not by caption
In September 1995 Myron Flickner and eleven other researchers at IBM Almaden described the QBIC system in IEEE Computer. An image and video database could be queried by example, by sketch, or by a chosen colour or texture rather than by words. Answers were ranked by similarity of colour, texture, shape and motion computed from the images themselves.
Why it matters
Until then an image in a database was found by the words someone had attached to it. QBIC made the image itself the query, and the likeness of two pictures a number the machine computes without a person. The paper’s abstract says the technology was already part of several IBM products.
Computer 28(9), pp. 23-32. At the SIGMOD ’95 demonstration in May of the same year the system ran on an RS/6000 under AIX, feature vectors had a dimension of about 64, a query could be combined with SQL and keyword predicates, and the test database held over 2300 images. A query from that demonstration: a green background with a round red object in the upper left. The record rests on the journal abstract and the full text of the demonstration: the article’s full text is paywalled and was not read. Neither the abstract nor the demonstration names the IBM products that carried the technology; the team itself named one in the title of a paper at Spring COMPCON ’94 in San Francisco (28 February-4 March 1994), "Ultimedia Manager: Query By Image Content and its applications", whose text was not read either, so the record does not say whether the product was already on sale. The project’s first paper, in the SPIE volume of the symposium held in San Jose from 31 January to 5 February 1993, was not read: the publisher’s page shows a CAPTCHA.