中央研究院 資訊科學研究所

Event detection is a long-standing problem and enables many applications such as activity recognition and facial expression analysis. In this talk I will cover two relatively unexplored perspectives in event detection---unsupervised Temporal Commonality Discovery (TCD) and domain transfer on facial expression analysis. A naive exhaustive search approach to solve the TCD problem has a computational complexity quadratic to the length of each sequence, making it impractical for regular-length sequences. Instead, we propose an efficient Branch and Bound (B&B) algorithm to tackle the TCD problem. Using the proposed bounds, the B&B algorithm can efficiently find the global optimal solution. Our algorithm is general and can be applied to any feature quantified into histograms. To the best of our knowledge, this is the  first work that addresses unsupervised discovery of common events in videos. The second topic involves domain transfer issues in facial expression analysis. Specifically, we introduce a personalized classifier, which we refer as a Selective Transfer Machine (STM), for attenuating person-specific biases. STM achieves this effect by simultaneously re-weighting the training samples that are most relevant to each test subject.