Diversity In Ensembles Of Codebooks For Visual Concept Detection

Visual codebooks generated by the quantization of local descriptors allows building effective feature vectors for image archives. Codebooks are usually constructed by clustering a subset of image descriptors from a set of training images. In this paper we investigate the effect of the combination of an ensemble of different codebooks, each codebook being created by using different pseudo-random techniques for subsampling the set of local descriptors. Despite the claims in the literature on the gain attained by combining different codebook representations, reported results on different visual detection tasks show that the diversity is quite small, thus allowing for modest improvement in performance w.r.t. the standard random subsampling procedure, and calling for further investigation on the use of ensemble approaches in this context.