Using video objects and relevance feedback in video retrieval

Video retrieval is mostly based on using text from dialogue and this remains the most significant component, despite progress in other aspects. One problem with this is when a searcher wants to locate video based on what is appearing in the video rather than what is being spoken about. Alternatives such as automatically-detected features and image-based keyframe matching can be used, though these still need further improvement in quality. One other modality for video retrieval is based on segmenting objects from video and allowing endusers to use these as part of querying. This uses similarity between query objects and objects from video, and in theory allows retrieval based on what is actually appearing on-screen. The main hurdles to greater use of this are the overhead of object segmentation on large amounts of video and the issue of whether we can actually achieve effective object-based retrieval. We describe a system to support object-based video retrieval where a user selects example video objects as part of the query. During a search a user builds up a set of these which are matched against objects previously segmented from a video library. This match is based on MPEG-7 Dominant Colour, Shape Compaction and Texture Browsing descriptors. We use a user-driven semi-automated segmentation process to segment the video archive which is very accurate and is faster than conventional video annotation.