中央研究院 資訊科學研究所

Information revolution is one of the most exciting revolutions in recent human history. In this talk, we will present how embracing diversity from various information assists two different problems, video summarization and event detection. We will first present a new perspective to video summarization, termed video co-summarization, that mines visual co-occurrence among diverse videos, and use the co-occurrences for summarizing a video set that shares the same topic. Our method is simple yet effective, and can be parallelized with closed-form updates. Our results suggest that summaries generated by visual co-occurrences are closer to human summarization compared to existing unsupervised video summarization methods. Second, we will introduce an ensemble-like method for event detection, termed Cascade of Tasks, which combines the use of different tasks (i.e., frame, segment and transition) for AU event detection. In addition, a new event-based metric is introduced to evaluate detection performance at segment-level. The effectiveness over state-of-the-art approaches is demonstrated in both frame-based and event-based metrics across three benchmark datasets that differ in complexity.

 

Wen-Sheng Chu is currently pursuing the Ph.D. degree at the Robotics Institute at Carnegie Mellon University. He received the B.S. and M.S. degrees in computer science from National Cheng Kung University, Tainan, Taiwan, in 2005 and 2007, respectively. In 2009 and 2010, he joined the Institute of Information Science at Academia Sinica, Taipei, Taiwan, and the Robotics Institute at Carnegie Mellon University, USA, as a research assistant. His research interests focus on computer vision and machine learning, especially automatic face analysis and unsupervised commonality discovery in spatial and temporal domains.