MRIM-LIG at ImageCLEF 2010 Visual Concept Detection and Annotation task

This paper focuses on one of the Image CLEF Photo tasks at which the MRIM research group of the LIG participated: the Visual Concept Detection and Annotation. For this task, we applied a simple state of the art technique based on bag of visual words. We extracted SIFT-like features that integrate colors (rgSIFT) proposed by van de Sande[10]. We used then a Kmeans clustering in a way to group these features according to 4000 clusters. We generated then for each image of the training set a 4000 dimensions histogram by summing all the occurrences of each cluster, using the nearest neighbour centroid for each extracted feature. For the recognition we extracted the rgSIFT features from the test set, before generating the 4000 dimensional histograms. We applied then SVMs with RBF kernels using a probabilistic estimation of recognition. The results obtained by our run are presented.