Eurecom at TRECVid 2007: Extraction of High-level Features

As past years we participated to the high-level feature extraction task and we pursued on the fusion of classifier outputs. New for this year is an experiment with a twodimensional Hidden Markov Model. Altogether we submitted seven runs. Three runs was based on the SVM model, another three was based on the HMM and one run was done by fusing HMM results with the SVM. To compile the runs, color and texture features were extracted from shot key-frames. Then, SVM and HMM classifiers were build per concept on the training data set. The fusion of classifier outputs is finally provided either by second level SVM or by hierarchical genetic fusion of possibilities (HGFP) on per concept basis. ”A RO1...1” fuses the output of both classifiers trained on color and texture features using HGFP. ”A RO1...2” fuses the output of SVM classifiers build on color and texture features using SVM. ”A RO1...4” fuses the output of SVM classifiers build on color and texture features using HGFP. ”A RO1...6” fuses using HGFP the output of SVM classifiers build on color and texture features and the output of an SVM trained on both features. The comparison of performances of the fusion systems shows that HGFP can efficiently fuse classifier outputs in a simple mannner. We also noticed that including the fusion at an earlier stage could improve retrieval performances. This year we took the opportunity to experiment with an early version of the implementation of a contextdependant classifier based on a two dimensional Hidden Markov Model. The HMM-model considers an image as a random process of observations. To provide the observations we divide the image into a grid of blocks where each block presents its color and frequency characteristics. We could observe the problem of a well known drawback of the HMMs, that the output probability plays a more important role than the transition probabilities.