Evaluations of multi-learner approaches for concept indexing in video documents

In this paper, we evaluated and compared multi-learner approaches for concept indexing in video documents. Multilearner approaches are designed to handle the problem of the sparse concepts that leads to a strong imbalance between the size of positive and negative sample sets. The evaluation and comparison have been carried out in the context of the concept detection task at TRECVID 2008 and 2009. The multi-learner method was experimented with three types of classifiers: SVM with linear and Gaussian kernels, and logistic regression. Methods were evaluated using several types of descriptors. The results were quite stable compared to the considered type of descriptor. Multi-learner methods do perform better than their mono-learner peers. In this context, logistic regression performs better than a linear SVM but less so than a SVM with a Gaussian kernel.