Today's search engines provide instant keyword-based auto-suggestion and completion of the user's search queries. This demo presents a novel auto-suggestion interface for the Semantic Multimedia Explorer (SEMEX), a semantic search engine that supports entity-based exploratory video retrieval. In difference to traditional text-based retrieval, auto-suggestion and auto-completion of the user's query string is not based on plain text but on semantic entities grouped by meaningful categories. Suggested entities are ranked by plain edit-distance as well as by popularity while their representations are enabled for brushing and linking. The categories contribute to a quick comprehensibility of the suggested entities compared to other systems that constrain an actually significant and practical feature into static and narrow vertical listings. Thus, our approach is leading to better decision making from the very start of an exploratory search.
Content Based Multimedia Retrieval on non-textual documents is constrained by available metadata. User-generated tags constitute an important source of information about a resource. To enable search scenarios exceeding traditional text-based search, such as exploratory and semantic search, this textual information must be complemented with semantic entities. Due to tag ambiguities and creative neologisms automatic semantic annotation based on user tags represents a major challenge. In this work, we show how to adopt context information and ontological knowledge to automatically assign semantic entities to user-generated tags for video data. Thus, a sophisticated semantic search on semantic entities is enabled. Our algorithm combines co-occurence and link graph analysis using Linked Data.
Text displayed in a video carries important information of video contents. Therefore, it provides a valuable source for indexing and retrieval of digital video libraries. In this paper, we propose a solution for automated text detection in video frames: Firstly, we developed an edge-based multi-scale text detector to detect the text coarsely. It serves to achive a high recall rate with low computational time expenses. Then, can- didate text lines are refined by an image entropy based refine- ment algorithm and a Stroke Width Transform (SWT) based verification procedure. Both, overlay text and recorded scene text can be localized. The accuracy of the proposed approach is proven by evaluation.
This demonstration presents the semantic search engine CONTENTUS and its user interface, a multimedia retrieval system developed in the context of the German research project THESEUS. This interface uses content-based suggestions and a multi-modal presentation of search results to support semantic search. In addition, the system deploys a combination of a facetted browsing and breadcrumb-based navigation; a time-line enables time based filtering of the search results and the system suggests related search results according to the users’ preferences. Finally, CONTENTUS has become more than the sum of its parts. Its unique feature combination facilitates search to become a more efficient and overall more pleasant user experience. This screencast demonstrates the capabilities of the CONTENTUS user interface by means of various search scenarios.
