Naming people in news videos with label propagation

Labeling persons appearing in video frames with names detected in a corresponding video transcript helps improving video content annotation and search tasks. We implement a face naming method that learns from labeled and unlabeled examples using iterative label propagation in a graph of connected faces or name-face pairs. By incorporating the unlabeled data points during the learning process, this method can work with few labeled data points. Moreover, we present variations of this model that better cope with a large number of data by reducing the time and space complexity. On BBC News videos, the label propagation algorithm yields better results than a Support Vector Machine classifier and a nearest neighbor classifier trained on the same labeled data. Furthermore, we show that when anchor detection precedes the label propagation process, it helps boosting the face naming performance. Reusing labeled examples from different broadcasts, we manage to name 70% of the faces with a precision of 85%.