Search

Learning image similarity plays a critical role in real-world multimedia information retrieval applications, especially in Content-Based Image Retrieval (CBIR) tasks, in which an accurate retrieval of visually similar objects largely relies on an effective image similarity function. Crafting a good similarity function is very challenging because visual contents of images are often represented as feature vectors in high-dimensional spaces, for example, via bag-of-words (BoW) representations, and traditional rigid similarity functions, for example, cosine similarity, are often suboptimal for CBIR tasks. In this article, we address this fundamental problem, that is, learning to optimize image similarity with sparse and high-dimensional representations from large-scale training data, and propose a novel scheme of Sparse Online Learning of Image Similarity (SOLIS). In contrast to many existing image-similarity learning algorithms that are designed to work with low-dimensional data, SOLIS is able to learn image similarity from large-scale image data in sparse and high-dimensional spaces. Our encouraging results showed that the proposed new technique achieves highly competitive accuracy as compared to the state-of-the-art approaches but enjoys significant advantages in computational efficiency, model sparsity, and retrieval scalability, making it more practical for real-world multimedia retrieval applications. [ABSTRACT FROM AUTHOR]

The prevailing of Web 2.0 techniques has led to the boom of web video content as well as its social network. To overcome the information overload problem, effective web video topic discovery and structuring techniques are highly demanded. To this end, existing works go to two respective directions: video topic discovery based on content or community detection in social network, with limited interplay between topics and network structures. In this paper, we construct the video social network based on web user interactions over videos. By comparing the topics and communities discovered on this network, we unveil the loose correspondence relationship between content and social network, and correspondingly propose a novel community-driven web video topic discovery model, which regularizes the topic model in relaxed community-level. Quantitatively analysis on real-world YouTube data shows that our model has achieved a significant improvement over the purely content-based or network-based baselines. Meanwhile, we propose a community-based topic structuralization framework, which decomposes a topic in social network space, and tracks the spreading trajectory of this topic among different communities on the time line. This structuralization can help users to catch the important facets of topics, such as “Who is interested with this topic” and “How does it propagate among the communities”, which provide valuable insights in related applications such as web monitoring and market development. [ABSTRACT FROM AUTHOR]

This paper studies the data redundancy of the coefficient matrix of the corresponding discrete system which forms a basis for fast algorithms of solving the integral equation whose kernel includes a convolution function factor. We develop lossless matrix compression strategies, which reduce the cost of integral evaluations and the storage to linear complexity, i.e., the same order of the approximation space dimensions. We establish that this algorithm preserves the convergence order of the approximate solution. We also propose a hardware-aware parallel algorithm for these strategies. [ABSTRACT FROM AUTHOR]

Abstract: Domain adaptive video concept detection and annotation has recently received significant attention, but in existing video adaptation processes, all the features are treated as one modality, while multi-modalities, the unique and important property of video data, is typically ignored. To fill this blank, we propose a novel approach, named multi-modality transfer based on multi-graph optimization (MMT-MGO) in this paper, which leverages multi-modality knowledge generalized by auxiliary classifiers in the source domains to assist multi-graph optimization (a graph-based semi-supervised learning method) in the target domain for video concept annotation. To our best knowledge, it is the first time to introduce multi-modality transfer into the field of domain adaptive video concept detection and annotation. Moreover, we propose an efficient incremental extension scheme to sequentially estimate a small batch of new emerging data without modifying the structure of multi-graph scheme. The proposed scheme can achieve a comparable accuracy with that of brand-new round optimization which combines these new data with the data corpus for the nearest round optimization, while the time for estimation has been reduced greatly. Extensive experiments over TRECVID2005-2007 data sets demonstrate the effectiveness of both the multi-modality transfer scheme and the incremental extension scheme. [Copyright &y& Elsevier]

Considering the challenges of using SVM to learn concepts from large-scale imbalanced datasets, we proposed a new method: Boosted Near-miss Under-sampling on SVM ensembles (BNU-SVMs). The BNU-SVMs is under the framework of under-sampling ensemble method, where a sequence of SVMs is trained and the training dataset for each base SVM is selected by a Boosted Near-miss Under-sampling technique. More specifically, by adaptively updating weights over negative examples, the most near-miss negative examples in output space are selected in each iteration. Since the training dataset is balanced and reduced by under-sampling and the performance of classifier is improved by ensembles, the BNU-SVMs is a promising solution for large-scale and imbalance problem. Moreover, the negative examples selected by BNU-SVMs not only contain the most representative ones from data distribution perspective, but also cover the easily misclassified ones from data accuracy perspective. Therefore, the outperformance of the BNU-SVMs is expected. In addition, considering the computation cost caused by high-dimensional visual features, we proposed a kernel-distance pre-computation technique to further improve the efficiency of the BNU-SVMs. Experiments on TRECVID benchmark datasets show that the BNU-SVMs outperforms the previous methods significantly, which demonstrates that the BNU-SVMs is a both effective and efficient solution to concept detection in large-scale imbalanced datasets. [ABSTRACT FROM AUTHOR]

Abstract: Topic models have been successfully used in information classification and retrieval. These models can capture word correlations in a collection of textual documents with a low-dimensional set of multinomial distribution, called “topics”. However, it is important but difficult to select the appropriate number of topics for a specific dataset. In this paper, we study the inherent connection between the best topic structure and the distances among topics in Latent Dirichlet allocation (LDA), and propose a method of adaptively selecting the best LDA model based on density. Experiments show that the proposed method can achieve performance matching the best of LDA without manually tuning the number of topics. [Copyright &y& Elsevier]