Incremental probabilistic Latent Semantic Analysis for video retrieval

a r t i c l e i n f o Recent research trends in Content-based Video Retrieval have shown topic models as an effective tool to deal with the semantic gap challenge. In this scenario, this paper has a dual target: (1) it is aimed at studying how the use of different topic models (pLSA, LDA and FSTM) affects video retrieval performance; (2) a novel incre-mental topic model (IpLSA) is presented in order to cope with incremental scenarios in an effective and efficient way. A comprehensive comparison among these four topic models using two different retrieval systems and two reference benchmarking video databases is provided. Experiments revealed that pLSA is the best model in sparse conditions, LDA tend to outperform the rest of the models in a dense space and IpLSA is able to work properly in both cases. With the expansion of new technologies, video collections are increasingly larger and more complex, therefore one of the biggest current challenges is how to retrieve users' relevant data from this huge amount of information. The Content-based Video Retrieval (CBVR) problem is concerned about how to provide users with videos which satisfy their queries by means of video content analysis. Over the past years, CBVR has become a very important research field and several CBVR systems have been developed [1–4]. In general, a CBVR system has three main components involved in the retrieval process: (1) a query, represented by a few video examples of the semantic concept that the user is looking for; (2) a database, which is used to extract videos related to the query concept; and (3) a ranking function, which sorts the database according to the relevance to the query. These three components are usually integrated together with the user in a Relevance Feedback (RF) scheme [5] to provide the most relevant videos through several feedback iterations. One of the most used rankings in multimedia retrieval is distance-based ranking. Such ranking is performed according to the minimum distance or maximum similarity to the query in the video representation space [6,7]. However, these measures tend not to work properly when the multimedia data is rather complicated [8]. Other ranking algorithms are based on inductive learning [9,10] which typically use a bank of classifiers to represent a set of possible events to test. Nevertheless, the performance of this approach heavily depends on the training data that limits its usage in …