Eurecom at TRECVid 2007: Extraction of High-level Features

In this paper we describe our experiments for the high level features extraction task of TRECVid 2007. Our approach is different than previous submissions in that we have implemented a multi-descriptors system. Five (5) experimentations are submitted based on: • Run 1: MPEG-7 global descriptors, • Run 2: MPEG-7 global and GET audio descriptors, • Run 3: MPEG-7 region descriptors using region based automatic segmentation method RBAS (A region merging approach incorporating geometric properties), • Run 4: Color and texture descriptors are extracted using three segmentation methods (A fixed image grid, watersheds and a technique based on minimum spanning trees MST), • Run 5: Combination of global and regions descriptors. The experimental results show that the performance can be improved with suitable concept models. Secondly, using audio features did not lead to performance improvement in our experiments. Keywords Video semantic analysis, multi-level fusion, feature fusion, c...

• Type: Conference
• Year: 2007
• Where: TRECVID
• Authors: Rachid Benmokhtar, Eric Galmar, Benoit Huet
• Keywords: Global Descriptors, MPEG-7, Region Descriptors, TRECVID 2007, TRECVID 2008