This paper describes the KU-ISPL Multimedia Event Detection (MED) system employed for competing in the TRECVID 2015 hosted by NIST. The proposed system utilizes diverse audio/visual information source components which consist of a combination of low-level and semantic-level features and adopts an optimized information fusion technique for accurate and robust video event detection. In particular, the local descriptors of the system are composed of the heterogeneous features extracted from the audio/visual information sources of video contents. The fusion process combines the information source at either low-level or semantic-level so that the individual detection score for video events can be judicially appraised in terms of meaningful representation and significance. The results from self-test and the official evaluation have indicated that the proposed system outperforms the previous version.
|Published - 2020
|2015 TREC Video Retrieval Evaluation, TRECVID 2015 - Gaithersburg, United States
Duration: 2015 Nov 16 → 2015 Nov 18
|2015 TREC Video Retrieval Evaluation, TRECVID 2015
|15/11/16 → 15/11/18
Bibliographical notePublisher Copyright:
© 2015 TRECVID.
ASJC Scopus subject areas
- Information Systems
- Signal Processing
- Electrical and Electronic Engineering