A new semantic visual features (e.g., car, mountain, and fire) navigation technology is proposed to improve the effectiveness of video retrieval. Traditional temporal neighbor browsing technology allows users to navigate temporal neighbors of a selected sample frame to find additional matches, while semantic visual feature browsing enables users to navigate keyframes that have similar features to the selected sample frame. A pilot evaluation was conducted to compare the effectiveness of three video retrieval designs that support 1) temporal neighbor browsing; 2) semantic visual feature browsing; and 3) fused browsing which is a combination of both temporal neighbor and semantic visual feature browsing. Two types of searching tasks: visual centric and non-visual centric tasks were applied. Initial results indicated that the semantic visual feature browsing system was more efficient for non-visual centric tasks.
IntroductionAccess to digital video from news sources such as CNN, MSNBC, or ABC has become commonplace. To make digital multimedia resource discovery and search more convenient, multimedia digital libraries are being developed for research and education.Increasingly, students or instructors are consulting video col lections in search of video shots within larger video "documents" to be used in their projects or lectures. Viewing all videos in full length to find the desired video shots may be feasible for a small collection, but can be very time intensive for a large collection. The ability to search within individual videos, much in the same way that full text searching allows users to search for content instead of their bibliographic surrogates, would g reatly increase access to video content. Recent research on content-based video retrieval indicated that initially performing a text-based query and subsequently proceeding with neighbor or visual similarity browsing proved to be an effective retrieval strategy (Wildemuth et al., 2003;Heesch et al., 2004;Mezaris et al., 2004 ; Amir et al., 2005). Human beings are usually good at pattern recognition through navigation. A retrieval system supporting navigation functions would provide users additional means for content rel ated searching tasks.In this paper we propose a new video content browsing techniqu e: semantic visual feature browsing. Our purpose is to evaluate its effectiveness as compared to traditional temporal neighbor browsing technique for two types of retrieval tasks: visual centric tasks and non-visual centric tasks. After the introduction of related research, a description of the semantic visual feature browsing algorithm will be given. The user interface of a prototype web-based video retrieval system that supports semantic visual feature browsing will be then illustrated. Finally, the methodology of a pilot user study and some initial results from the study will be presented, fol lowed by a brief discussion.
Related ResearchVideo retrieval in the context of a digital library has only recently begun to be studied from a research perspective...