Automatic video annotation with forests of fuzzy decision trees

Abstract

Nowadays, the annotation of videos with high-level semantic concepts or features is a great challenge. In this paper, this problem is tackled by learning, by means of Fuzzy Decision Trees (FDT), automatic rules based on a limited set of examples. Rules intended, in an exploitation step, to reduce the need of human usage in the process of indexation. However, when addressing large, unbalanced, multiclass example sets, a single classi er - such as the FDT - is insu cient. Therefore we introduce the use of forests of fuzzy decision trees (FFDT) and we highlight: (a) its e ectiveness on a high level feature detection task, compared to other competitive systems and (b) the e ect on performance from the number of classi ers point of view. Moreover, since the resulting indexes are, by their nature, to be used in a retrieval application, we discuss the results under the lights of a ranking (vs. a classi cation) context.