Vireo-Tno @ TRECVID 2014: Multimedia event detection and recounting (MED and MER)

• Main Authors: NGO, Chong-wah, LU, Yi-Jie, ZHANG, Hao, YAO, Ting, TAN, Chun-Chet, PANG, Lei, DE BOER, Maaike, SCHAVEMAKER, John, SCHUTTE, Klamer, KRAAIJ, Wessel
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2014
• Subjects: Computer Sciences; Graphics and Human Computer Interfaces
• Online Access: https://ink.library.smu.edu.sg/sis_research/6646; https://ink.library.smu.edu.sg/context/sis_research/article/7649/viewcontent/med_and_mer.pdf
• Institution: Singapore Management University

This paper presents an overview and comparative analysis of our systems designed for TRECVID 2014 [1] multimedia event detection (MED) and recounting (MER) tasks, including all sub-tasks for Pre-Specified (PS) event detection, all sub-tasks except 100Ex for Ad-Hoc (AH) event detection, and 010Ex sub-task for both PS and AH event recounting. Multimedia Event Detection (MED) : Our main focus for the MED task is on the study of a new zero-example system, which aims to solve the 000Ex and SQ problems. The system can run either fully automatically or semi-automatically. Specifically, we test the automatic run in 000Ex submission and the semi-automatic run in SQ submission. Multimedia Event Recounting (MER) : For the 010Ex sub-tasks of both PS and AH event recounting, we design and implement a simple but effective system to optimize the concept-to-event relevance, evidence diversity and timing of evidential shots. We show that good performance can be achieved by only selecting the three shots with the highest confidence.