New Benchmark Datasets and a Character Identification System on TV Series

Automatic character identification has been extensively studied in the past decades. However, most related works find characters only depending on their faces. One important reason is no dataset containing labels with all the occurrences of characters. Towards this end, we propose 3 datasets for character identification, which consist of 15 episodes (573 minutes). We label all the frames in which specified characters can be identified manually, regardless of whether faces or persons are detected. To the best of our knowledge, they are the first and largest datasets with all the occurrences of specified characters labelled for automatic character identification on TV material. Based on these datasets, we propose an automatic character identification system. Given a query image with a clear frontal face of a character, we can effectively utilize identified frames as new queries to explore remaining ones. In experiments, we show the average precision is substantially boosted in many cases.