Temporal Relation Analysis in Audiovisual Documents for Complementary Descriptive Information

摘要

Relations among temporal intervals can be used to detect semantic events in audio visual documents. The aim of our work is to study all the relations that can be observed between different segmentations of a same document. These segmentations are automatically provided by a set of tools. Each tool determines temporal units according to specific low or mid-level features. All this work is achieved without any prior information about the document type (sport, news...), its structure, or the type of the semantic events we are looking for. Considering binary temporal relations between each couple of segmentations, a parametric representation is proposed. Using this representation, observations are made about temporal relation frequencies. When using relevant segmentations, some semantic events can be inferred from these observations. If they are not relevant enough, or if we are looking for semantic events of a higher level, conjunctions between two temporal relations can turn to be more efficient. In order to illustrate how observations can be made in the parametric representation, an example is given using Allen's relations. Finally, we present some first results of an experimental phase made on TV news programs.