Image + database ≠ image database

For decades, computer vision researchers have been trying to extract high level information from images. While the semantics of images is still unreachable from the signal in most real cases, users would like to express requests to image data bases using high-level queries. This gap between the user needs and the image processing capabilities will limit the use of image databases in the near to mid term future.Meaningful applications, however, are already possible using existing scientific technology, for instance using query-by-example. The scalability of such applications stresses the need for: new indexing methods able to handle approximate measures from the image signal; approximate search methods that are efficient in high dimensional spaces; and robust search methods able to handle many partially erroneous data (outliers).The tutorial will illustrate some limited answers to these open problems using invariant features, robust statistics and probabilistic matching. It will then focus on the long term goal of high level semantics extraction from images. This problem is as yet poorly defined: the semantics of an image is user dependent and nobody knows how to express it in a formal way. Some limited answers exist, however, and the tutorial will illustrate how learning mechanisms provide impressive initial results. Moreover learning can be linked to relevance feedback and therefore allows performing better user dependent search.