The France Telecom Orange Labs ( Beijing ) Video High - level Feature Extraction Systems – TrecVid 2009 Notebook Paper

In this paper, we described the video high-level feature extraction systems developed at France Telecom Orange Labs (Beijing). In our systems, four categories of lowlevel visual features, namely color, edge, texture and SIFT local descriptors, were extracted. Two approaches to fusing the representative capabilities of these visual features were investigated for different runs. Under the setting of late fusion, separate SVM classifiers were constructed for each low-level visual feature and their outputs were combined in a weighted manner. While under the setting of early fusion, a composite kernel was constructed to merge multiple kernels of various visual features and one SVM was trained on such composite kernel. The evaluation results on the 2009 TrecVid highlevel feature extraction task were presented, among which the A_FTRD-HLF-5 run achieved an MAP of 0.17 and was above the median for all concepts.