News Monologue Shot Detection using Conditional Random Fields

In TV news videos, monologue shots are informative and valuable in the application of video retrieval and mining. In this paper, we employ conditional random fields (CRFs) to fuse contextual information as well as audio, visual and temporal features for the detection of news monologue shots. CRFs are undirected probabilistic models and deal with monologue shot detection as a sequence labeling problem. The method first removes commercial shots, and then applies a two-level framework to detect monologue shots. In the low-level model, a face detector and an anchorperson detector are employed to identify the corresponding shots. In the high-level model, monologue and reporter shots are labeled with CRFs. The experimental results achieve better performance without external knowledge.