c4851e8e264415c4094e4e85b0baa7cc-Reviews.html

This paper considers automatic classification of unstructured social group activity videos. To bridge the semantic gap between low-level features and the class-labels, the authors adopt a latent topic model based on replicated softmax to extract topics as mid-level representations for video classification. The main idea of this paper is the integration of sparse Bayesian learning and replicated softmax, which leads to the proposed model referred to "relevance topic model (RTM)". In RTM, the discriminative topics and sparse classifier weights are learned jointly, and the authors proposes variational EM algorithm for model parameter estimation and inference. The authors test their algorithm on a benchmark dataset and demonstrate better performance compared to other supervised topic models and some baseline algorithms.