Search

In this paper we describe our system and experiments performed for both the automatic search task and the event detection task in TRECVid 2008. For the automatic search task for 2008 we submitted 3 runs utilizing only visual retrieval experts, continuing our previous work in examining techniques for query-time weight generation for data-fusion and determining what we can get from global visual only experts. For the event detection task we submitted results for 5 required events (ElevatorNoEntry, OpposingFlow, PeopleMeet, Embrace and PersonRuns) and 1 optional event (DoorOpenClose).

In this paper we describe our retrieval system and experiments performed for the automatic search task in TRECVid 2007. We submitted the following six automatic runs: • F A 1 DCU-TextOnly6: Baseline run using only ASR/MT text features. • F A 1 DCU-ImgBaseline4: Baseline visual expert only run, no ASR/MT used. Made use of query-time generation of retrieval expert coefficients for fusion. • F A 2 DCU-ImgOnlyEnt5: Automatic generation of retrieval expert coefficients for fusion at index time. • F A 2 DCU-imgOnlyEntHigh3: Combination of coefficient generation which combined the coefficients generated by the query-time approach, and the index-time approach, with greater weight given to the index-time coefficient. • F A 2 DCU-imgOnlyEntAuto2: As above, except that greater weight is given to the query-time coefficient that was generated. • F A 2 DCU-autoMixed1: Query-time expert coefficient generation that used both visual and text experts.

For the 2006 Terabyte track in TREC, Dublin City University’s participation was focussed on the ad hoc search task. As per the pervious two years [7, 4], our experiments on the Terabyte track have concentrated on the evaluation of a sorted inverted index, the aim of which is to sort the postings within each posting list in such a way, that allows only a limited number of postings to be processed from each list, while at the same time minimising the loss of effectiveness in terms of query precision. This is done using the Físréal search system, developed at Dublin City University [4, 8].

In this paper we describe our retrieval system and experiments performed for the automatic search task in TRECVid 2006. We submitted the following six automatic runs: • F A 1 DCU-Base 6: Baseline run using only ASR/MT text features. • F A 2 DCU-TextVisual 2: Run using text and visual features. • F A 2 DCU-TextVisMotion 5: Run using text, visual, and motion features. • F B 2 DCU-Visual-LSCOM 3: Text and visual features combined with concept detectors. • F B 2 DCU-LSCOM-Filters 4: Text, visual, and motion features with concept detectors. • F B 2 DCU-LSCOM-2 1: Text, visual, motion, and concept detectors with negative concepts. The experiments were designed both to study the addition of motion features and separately constructed models for semantic concepts, to runs using only textual and visual features, as well as to establish a baseline for the manually-assisted search runs performed within the collaborative K-Space project and described in the corresponding TRECVid 2006 notebook paper. The results of the experiments indicate that the performance of automatic search can be improved with suitable concept models. This, however, is very topic-dependent and the questions of when to include such models and which concept models should be included, remain unanswered. Secondly, using motion features did not lead to performance improvement in our experiments. Finally, it was observed that our text features, despite displaying a rather poor performance overall, may still be useful even for generic search topics.

For the 2005 Terabyte track in TREC Dublin City University participated in all three tasks: Adhoc, E±ciency and Named Page Finding. Our runs for TREC in all tasks were primarily focussed on the application of "Top Subset Retrieval" to the Terabyte Track. This retrieval utilises different types of sorted inverted indices so that less documents are processed in order to reduce query times, and is done so in a way that minimises loss of effectiveness in terms of query precision. We also compare a distributed version of our Físréal search system [1][2] against the same system deployed on a single machine.

In this paper we describe our experiments in the automatic and interactive search tasks and the BBC rushes pilot task of TRECVid 2005. Our approach this year is somewhat different than previous submissions in that we have implemented a multi-user search system using a DiamondTouch tabletop device from Mitsubishi Electric Research Labs (MERL).We developed two versions of oursystem one with emphasis on efficient completion of the search task (Físchlár-DT Efficiency) and the other with more emphasis on increasing awareness among searchers (Físchlár-DT Awareness). We supplemented these runs with a further two runs one for each of the two systems, in which we augmented the initial results with results from an automatic run. In addition to these interactive submissions we also submitted three fully automatic runs. We also took part in the BBC rushes pilot task where we indexed the video by semi-automatic segmentation of objects appearing in the video and our search/browsing system allows full keyframe and/or object-based searching. In the interactive search experiments we found that the awareness system outperformed the efficiency system. We also found that supplementing the interactive results with results of an automatic run improves both the Mean Average Precision and Recall values for both system variants. Our results suggest that providing awareness cues in a collaborative search setting improves retrieval performance. We also learned that multi-user searching is a viable alternative to the traditional single searcher paradigm, provided the system is designed to effectively support collaboration.

In TREC2004, Dublin City University took part in three tracks, Terabyte (in collaboration with University College Dublin), Genomic and Novelty. In this paper we will discuss each track separately and present separate conclusions from this work. In addition, we present a general description of a text retrieval engine that we have developed in the last year to support our experiments into large scale, distributed information retrieval, which underlies all of the track experiments described in this document.

TREC-2002 saw the second running of the Video Track, the goal of which was to promote progress in content-based retrieval from digital video via open, metrics-based evaluation. The track used 73.3 hours of publicly available digital video (in MPEG-1/VCD format) downloaded by the participants directly from the Internet Archive (Prelinger Archives) (internetarchive, 2002) and some from the Open Video Project (Marchionini, 2001). The material comprised advertising, educational, industrial, and amateur films produced between the 1930's and the 1970's by corporations, nonprofit organizations, trade associations, community and interest groups, educational institutions, and individuals. 17 teams representing 5 companies and 12 universities - 4 from Asia, 9 from Europe, and 4 from the US - participated in one or more of three tasks in the 2001 video track: shot boundary determination, feature extraction, and search (manual or interactive). Results were scored by NIST using manually created truth data for shot boundary determination and manual assessment of feature extraction and search results. This paper is an introduction to, and an overview of, the track framework - the tasks, data, and measures - the approaches taken by the participating groups, the results, and issues regrading the evaluation. For detailed information about the approaches and results, the reader should see the various site reports in the final workshop proceedings.