University of Paris 6 at TRECVID 2006: Forests of Fuzzy Decision Trees for High-Level Feature Extraction

In this paper, we present the methodology we submitted to the NIST TRECVID'2006 evaluation. We participated in the High-level Feature Extraction task. Our approach is based on the use of a Forest of Fuzzy Decision Trees through the Salammbô software. 1 Structured Abstract Summary Here we present the contribution of the University of Paris 6 at TRECVID 2006 [1]. It concerns only the High-Level Feature Extraction task. The approach focuses on the use of a Forest of Fuzzy Decision Trees (FFDT) and is based on a rather simple image description. In the following, we start with a short summary of the used method and starting from Section 3, our approach is detailed. First, we describe the particularities of our image descriptors. Then we explain how the training (Section 4) and classi cation (Section 5) was performed. Before concluding, the submitted runs are discussed in details (Section 6). 1.1 Brief Description of the Submitted Run Here is the general information about the submitted run: The task: High-Level Feature Extraction. Type: A system trained on TRECVID development collection data, and common annotation of such data. Data used: XML les that provide the time-codes of each shot (Master shot references by [6]), All the image les (the keyframes), Annotations les for the devel keyframes. Pre-treatment: Each keyframe was segmented into 5 regions (see Section 3.1), An HSV histogram was computed for each region (see Section 3.1), Temporal information about each shot was extracted from the XML les. (see Section 3.2). Training: A Forest of Fuzzy Decision Trees (FFDT) was constructed and trained from the Devel data set (see Section 4). Ranking: The Forest of Fuzzy Decision Trees (FFDT) was used to rank the shots from the Test data set (see Section 5.2).