Home >

news Help

Publication Information

English:Velocity Pyramid for Event Detection 
Japanese: 梁卓琳, 井上中順, 篠田浩一.  
English: Zhuolin Liang, Nakamasa Inoue, Koichi Shinoda.  
Language English 
Journal/Book name
Japanese:電子情報通信学会技術研究報告 PRMU 
English:Technical Reports of IEICE PRMU 
Volume, Number, Page vol. 113    no. 493    pp. 13-18
Published date Mar. 13, 2014 
Conference name
English:Pattern Recognition and Media Understanding (PRMU) 
Conference site
Abstract In this paper, we propose a new motion feature, a velocity pyramid, for multimedia event detection. In an event which is a complex human activity, motion information is an important cue. However, most of the conventional motion features are too expensive when applied to event detection. Spatial pyramid matching introduces coarse geometric information into the Bag of Features framework. A velocity pyramid, which is motivated from spatial pyramid, can represent rough dynamic information. The idea is to categorize densely sampled features according to their velocity direction. It is effective for detecting events characterized by their temporal patterns. Experiments on the MED (Multimedia Event Detection) task of the TRECVID workshop have shown 20% improvement of the performance by velocity pyramid. Further, when combined with spatial pyramid, velocity pyramid provided an extra 5% gains to the detection performance. Also when compared with other motion features, the computation cost is reduced while keeping the performance.

©2007 Tokyo Institute of Technology All rights reserved.