In this paper we are interested in classification of objects in images according to user defined scenarios. We show how the user-defined ontology with a specialisation by a concrete scenario / object of interest allows for an adapted choice of methods and their tuning through the whole framework: selection of the area of interest, descriptors choice, classification of objects. Particular attention here is payed to the classification. We use SVM classifiers for their good capacity of generalisation. We show that in an adapted descriptor space, the choice of a “light” linear kernel together with boosting of classifiers is interesting compared to more complex and computationally expensive RBF kernels. The results on real-life images are promising. The paper results from the research we conduct in the framework of X-Media EU-funded Integrated Project.
展开▼
