University of Siegen, Kobe University and Muroran Institute of Technology at TRECVID 2013 Multimedia Event Detection
Kimiaki Shirahama, Chen Li, Marcin Grzegorzek, Kuniaki Uehara · TRECVID · 2013
This paper presents our method developed for TRECVID 2013 Multimedia Event Detection task. The following two problems are mainly addressed: The first is weakly supervised setting where training videos contain many shots irrelevant to a target event. The other is the diversity of visual appearances, meaning that shots relevant to the event are characterised by significantly different visual appearances. To overcome these problems, Hidden Conditional Random Fields (HCRFs) are used where the event is detected by assigning shots in a video to hidden states, each of which represents the compatibility between a feature combination and the event label. Although our submitted run SiegenKobeMuro MED13 VisualSys PROGAll PS 100Ex 6 on the progress search set was ranked at the low position (MAP of 4:1%), preliminary experiments on MED Test Background set show the effectiveness and potential of our method.