Effective and Efficient Indexing for Large Video Databases

Abstract

Content based multimedia retrieval is an important topic in database systems. An emerging and challenging topic in this area is the content based search in video data. A video clip can be considered as a sequence of images or frames. Since this representation is too complex to facilitate efficient video retrieval, a video clip is often summarized by a more concise feature representation. In this paper, we transform a video clip into a set of probabilistic feature vectors (pfvs). In our case, a pfv corresponds to a Gaussian in the feature space of frames. We demonstrate that this representation is well suited for accurate video retrieval. The use of pfvs allows us to calculate confidence values for frames or sets of frames for being contained within a given video in the database. These confidence values can be employed to specify two types of queries. The first type of query retrieves the videos stored in the database which contain a given set of frames with a probability that is larger than a given thresh-old value. Furthermore, we introduce a probabilistic ranking query retrieving the k database videos which contain the given query set with the highest probabilities. To efficiently process these queries, we introduce query algorithms on set-valued objects. Our solution is based on the Gauss-tree, an index structure for efficiently managing Gaussians in arbitrary vector spaces. Our experimental evaluation demonstrates that sets of probabilistic feature vectors yield a compact and descriptive representation of video clips. Additionally, we show that our new query algorithms outperform competitive approaches when answering the given types of queries on a database of over 900 real world video clips.