Citation Context

...elevance. This is explained by the fact that combining multimodal information requires finding a common and relevantsfeature space (for instance with information theory based criteriasas explained in =-=[14]-=-) to efficiently exploit the richness of eachsmodality which is not the case by simply added our two similaritysmatrix. However, combining the two cluster sets in uniquesclustering raises another impo...

Citation Context

...[23], the authors showed that this superiority also holds outside the space of learned semantic concepts, using multiple image queries. Another approach to QBSE, using the semantic labeling system of =-=[19]-=-, is presented in [18]. Although laying the foundations for QBSE, these previous works lack a systematic study of the QBSE paradigm. In this work, using the QBSE implementation of [23], we address thi...

Citation Context

... [12], and MBRM [10]; inference networks that connect image segments with words [24]; nonparametric density estimation [34]; supervised learning models [3, 4]; information-theoretic semantic indexing =-=[21]-=-; and [22] show that a proper selection of features could lead to very good results for a k-nearest neighbours algorithm. The human understanding of a scene was a topic confronted by many researchers ...

Citation Context

... vocabulary of visual terms for each keyword, Yavlinsky et al. [22] deployed a nonparametric distribution, Carneiro and Vasconcelos [1] a semi-parametric density estimation, while Magalhães and Rüger =-=[15]-=- engaged a maximum entropy framework. The above methods use features extracted from the multimedia itself, but heuristic techniques rely on metadata attached to the multimedia: for example, Lu et al [...

Citation Context

...of research: Snoek et al. [24] explore temporal synchronization to combine the multi-modal patterns, Monay and Gatica-Perez explore dependencies across different media [17], while Magalhães and Rüger =-=[15]-=- developed a multimodal maximum entropy framework. The above methods extract features from the multimedia itself, but other, heuristic techniques rely on metadata attached to the multimedia: for examp...

Citation Context

...ugh a voting scheme to produce a single classifier with better accuracy. More recent approaches concatenate the single-medium feature vectors into a single cross-media vector, see Magalhães and Rüger =-=[17]-=-. Although we also concatenate single-medium features into a single feature vector, our approach differs from the previous ones because we construct the cross-media feature vector by also taking into ...

Citation Context

...atical transformations are employed to determine a space where features extracted from heterogeneous sources (visual, audio, text) can be optimally combined. The work presented by Magalhães and Rüger =-=[2]-=- is an indicative example of this category where information theory and a maximum entropy model are utilized to integrate heterogeneous data into a unique feature space. In [3] Wu et al. work on the s...

Citation Context

...by hierarchically modeling their distributions of words and image feature descriptors. Grosky et al. [15] use Latent Semantic Indexing (LSI) and word weighting schemes to reduce the dimensionality. Feature vectors of visual feature descriptors and category label bits are concatenated in order to retrieve images. Benitez et al. [3] extract knowledge from annotated image collections by clustering the images based on visual 3 feature descriptors and text feature descriptors. Perceptual relationships, based on descriptor similarity and statistics between clusters, are discovered. Magalhaes et al. [25] use information theory to develop a model that supports heterogeneous types of documents (text documents, image documents, or documents with both text and images). Also, to take into account the subjectivity of human perception and bridge the gap between the high-level concepts and the low-level features, relevance feedback has been proposed to enhance the retrieval performance [20]. Finally, more sophisticated graphical models, such as the approach described in [5], based on Multinomial Dirichlet Mixture models or Gaussian-multinomial mixture model (GM-Mixture), Latent Dirichlet Allocator (L...