TRECVID is an independent evaluation of a video “track” in the TREC conference series, which is an annual worldwide competitive event since the year of 2000 and sponsored by NIST (National Institute of Standards and Technology) in the US with additional support from other US government agencies. The goal of TRECVID is to evaluate and encourage technology development on automatic segmentation, indexing and content-based retrieval of digital videos, in which three main tasks are designed, i.e. shot boundary detection (SBD), high-level feature extraction and search.

In 2007, LIVE consortium has participated in this well-known activity arranged by the coordinator Fraunhofer Institute IAIS, Germany, in which University of Bradford is delegated for the task of shot boundary detection (SBD). SBD is the fundamental task in cutting videos into natural shots by computerized means, which plays essential roles in video content processing, high-level feature extraction, and visual content editing and management. As the first time participant, Bradford researchers have submitted 9 runs on SBD as required by the TRECVID organizer. The technology developed by Bradford researchers is capable of shot boundary detections in compressed domain, where content features are directly extracted from DCT coefficients rather than pixel data, including block-based DC parameters (Y,Cb,Cr), motion information as well as energy and active macroblocks. Based on these features in compressed domain, detection of shot boundaries, including abrupt changes and gradual transitions like fade, dissolve and others such as wipe or combined shot changes, are characterized by combinational approaches based on correlation rules and finite state machines. Due to the compressed domain operation, the software developed by Bradford researchers achieves exceptionally fast speed and low computing cost, suitable for real-time applications.

Evaluation of all submissions is measured in terms of three categories, which are recall and precision rates of cut detection, gradual transition, and overall performances (consideration for detection of both cut and gradual transitions). For gradual transition, one additional parameter is assessed on frame accuracy in locating shot boundaries. According to the final results announced by NIST, the organizer of TRECVID, evaluation of Bradford’s submission can be highlighted as follows:

• For cut detection, Bradford submission is ranked No. 1 followed by Tsinghua University (joint submission with Intel Chinese Research Centre) from China and AT&T from US. The recall and precision rates achieved by Bradford is 97.3% and 98.2%, respectively;
• For gradual transition detection, Bradford submission is ranked No. 6, and the first two groups of best results are achieved by AT&T and Tsinghua/Intel Chinese Research Centre;
• Bradford submission is ranked No. 3 in overall evaluation, which has a recall rate of 94.1% and precision rate of 91.9% in comparison with (95.6%, 95.4%) from AT&T and (94.9%, 95.6%) from Tsinghua/Intel Chinese Research Centre;
• Owing to compressed-domain processing, Bradford submission achieves a speed of about 123 frames per second on our machine (PentiumD 2.8G/1G memory), including both partially decoding and segmentation, which is nearly 5 times faster than real-time playing of the video.

In 2007, the TRECVID test data set is different from previous years, covering a wide range of videos from news reports, documentaries, and educational programmes to archived videos in black and white. For SBD, about 6 hours of test data in MPEG-1 is selected from 400 hours video sources provided by the Netherlands Institute for Sound and Vision (http://portal.beeldengeluid.nl/). Among all the test videos, there are more than 2320 shots in 17 test sequences with 637,805 frames, in which about 90% of the shots are cut. For quantitative evaluation, ground truth data is manually extracted. It is worth noting that, even after hard working, there are still some errors in these ground truth data such as missing or false definitions, and this is mainly due to unclear boundary of some special editing effects and massive labors involved. In 2007, there are 35 teams registered for the SBD task, and eventually only 15 teams have made their submissions.

As the first attempt in attending this world-wide competitive event, Bradford team (Prof. Jianmin Jiang, Dr. Jinchang Ren and Miss Juan Chen) has successfully proved that LIVE consortium has great competence and potential for delivering high quality research work. Among the top three teams, Bradford is the only one from Europe, and the other two teams, one from China and the other from US, have been participating in the TRECVID event for the past three or four years.