A Language for Audiovisual Template Specification and Recognition

We address the issue of detecting automatically occurrences of high level patterns in audiovisual documents. These patterns correspond to recurring sequences of shots, which are considered as first class entities by documentalists, and used for annotation and retrieval. We introduce a language for specifying these patterns, based on an extension of Allen's algebra with the regular expression operator +, which denotes an iteration of arbitrary length. We propose a formulation of this pattern language using the constraint satisfaction framework, in which templates are represented as constraint problems. We propose an efficient representation of domains (all subsequences of a given graph) and filtering methods for the Allen constraints. We illustrate the resulting system on a corpus of real world news broadcast examples.